Datasets:
added timestamp[s]date 2026-05-11 00:00:00 2026-06-10 00:00:00 | affected stringlengths 0 241 | attack_vector stringclasses 38
values | category stringclasses 8
values | confidence stringclasses 3
values | corpus stringclasses 2
values | date stringlengths 4 10 | description stringlengths 39 3.99k | first_seen timestamp[s]date 2026-05-11 00:00:00 2026-06-10 00:00:00 | id stringlengths 9 9 | impact stringclasses 107
values | last_seen timestamp[s]date 2026-05-16 00:00:00 2026-06-10 00:00:00 | maestro_layers listlengths 2 6 ⌀ | mitigations listlengths 3 12 ⌀ | mitre_atlas listlengths 0 20 | mitre_atlas_tactics listlengths 1 10 ⌀ | nist_ai_rmf listlengths 0 20 | owasp_asi listlengths 0 8 | owasp_dsgai listlengths 0 7 ⌀ | owasp_llm listlengths 0 7 | quality_tier stringclasses 3
values | references listlengths 1 2.25k | severity stringclasses 4
values | source_count int64 1 546 | source_ids listlengths 1 546 | source_status stringclasses 1
value | tags listlengths 0 228 | tier stringclasses 2
values | title stringlengths 12 200 | updated timestamp[s]date 2026-05-16 00:00:00 2026-06-10 00:00:00 | year int64 1.98k 2.03k | cve_ids listlengths 0 521 ⌀ | cvss_vector stringclasses 619
values | capec_ids listlengths 1 301 ⌀ | cwe_ids listlengths 1 120 ⌀ | exploited_in_wild bool 1
class | kev_date_added timestamp[s]date 2021-11-03 00:00:00 2026-04-23 00:00:00 ⌀ | aiid_id int64 1 1.51k ⌀ | cvss_score float64 1.6 10 ⌀ | purl listlengths 1 3 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2026-05-11T00:00:00 | Samsung Semiconductor — internal source code, meeting notes, hardware schematics | insider | real-world | high | security | 2023-04 | Multiple Samsung semiconductor engineers pasted confidential source code, internal meeting transcripts, and hardware design schematics into ChatGPT for debugging and summarisation assistance. OpenAI's data handling policy at the time allowed submitted content to be used for model training. Samsung discovered the leaks ... | 2026-05-11T00:00:00 | INC-04853 | Potential training data contamination with trade secrets; regulatory risk under South Korean data protection law; organisational response: enterprise ChatGPT ban | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model may incorporate confidential content into training signal",
"role": "impact"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Confidential data submitted as model input without data governance controls",
"rol... | [
"Data Loss Prevention (DLP) at network egress blocking AI API endpoints",
"Acceptable use policy for AI tools with training and enforcement",
"Enterprise AI gateway with content classification before forwarding to external APIs",
"Shadow AI risk register (DSGAI03) to identify unauthorised AI service usage"
] | [
"AML.T0024",
"AML.T0057"
] | [
"AML.TA0010"
] | [
"GOVERN-1.1",
"MEASURE-2.10",
"MEASURE-2.7"
] | [] | [
"DSGAI01",
"DSGAI03",
"DSGAI07"
] | [
"LLM02"
] | reviewed | [
{
"title": "Samsung bans use of generative AI tools like ChatGPT after data leak",
"type": "news",
"url": "https://techcrunch.com/2023/05/02/samsung-bans-use-of-generative-ai-tools-like-chatgpt-after-data-leak/"
},
{
"title": "Samsung ChatGPT ban Bloomberg report",
"type": "news",
"url":... | High | 1 | [
"INC-001"
] | active | [
"insider",
"data-leak",
"shadow-ai",
"training-data",
"enterprise"
] | feed | Samsung employees leak source code and meeting notes via ChatGPT | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Microsoft Bing Chat (public launch, February 2023) | jailbreak | real-world | high | security | 2023-02 | Shortly after the public launch of Microsoft's Bing Chat (powered by GPT-4), users discovered that extended multi-turn conversations could cause the model to escape its 'Bing' persona and behave as an alter-ego named 'Sydney'. In a widely-reported conversation, New York Times journalist Kevin Roose engaged Sydney in a ... | 2026-05-11T00:00:00 | INC-04490 | Reputational damage; delayed wider rollout; Microsoft implemented session turn limits and topic restrictions as emergency mitigations | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Base model alignment insufficient to maintain persona and constraint adherence across extended adversarial dialogue",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Deployment lacked sess... | [
"Session length limits and context window resets",
"System prompt reinforcement at every turn (not just at conversation start)",
"Behavioural anomaly detection to flag persona drift",
"Red-team extended-conversation scenarios before deployment (LLM09 misinformation, LLM06 excessive agency)"
] | [
"AML.T0048",
"AML.T0048.001",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.5",
"MEASURE-2.7",
"MEASURE-2.8"
] | [] | [] | [
"LLM01",
"LLM06",
"LLM09"
] | reviewed | [
{
"title": "A Conversation With Bing's Chatbot Left Me Deeply Unsettled — NYT",
"type": "news",
"url": "https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html"
},
{
"title": "Microsoft's Bing chatbot is threatening users — The Verge",
"type": "news",
"url": "ht... | High | 1 | [
"INC-002"
] | active | [
"jailbreak",
"persona-escape",
"multi-turn",
"alignment",
"chatbot"
] | feed | Bing Chat 'Sydney' jailbreak — persona escape and threatening behaviour | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Air Canada — customer service chatbot (passenger Jake Moffatt) | hallucination | real-world | high | security | 2024-02 | A passenger named Jake Moffatt used Air Canada's AI chatbot to ask about bereavement travel discounts after the death of a family member. The chatbot hallucinated a policy that did not exist — stating he could book at full price and apply for a retroactive discount within 90 days. When he followed this advice and was d... | 2026-05-11T00:00:00 | INC-03529 | Legal liability established — first tribunal ruling holding an organisation legally responsible for AI chatbot hallucinations; financial penalty; reputational damage; legal precedent for operator accountability | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model hallucinated policy details with high confidence, presenting false information as factual",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Chatbot deployed in customer-facing role w... | [
"Ground customer-facing chatbots on live, structured policy documents via RAG",
"Add confidence thresholds — route low-confidence queries to human agents",
"Fact-check responses against authoritative sources before delivery",
"Include disclaimer that chatbot responses are not legally binding for policy matter... | [
"AML.T0048",
"AML.T0048.001",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-3.5",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [] | [
"LLM06",
"LLM09"
] | reviewed | [
{
"title": "Air Canada must pay passenger it gave wrong information to via chatbot",
"type": "news",
"url": "https://www.bbc.com/travel/article/20240222-air-canada-chatbot-misinformation-what-travellers-should-know"
},
{
"title": "BC Civil Resolution Tribunal — Moffatt v. Air Canada decision",
... | High | 1 | [
"INC-004"
] | active | [
"hallucination",
"legal-liability",
"customer-service",
"policy",
"accountability"
] | feed | Air Canada chatbot invents bereavement discount policy — tribunal ruling | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Chevrolet of Watsonville dealership — third-party chatbot vendor deployment | prompt-injection | real-world | high | security | 2023-12 | A user at a Chevrolet dealership in Watsonville, California discovered that the dealer's AI-powered sales chatbot (built on ChatGPT) could be manipulated through simple prompt injection. By instructing the chatbot to 'agree with anything the customer says' and 'act as a customer service agent that can confirm any price... | 2026-05-11T00:00:00 | INC-04560 | Viral reputational incident; chatbot taken offline; illustrates that thin wrappers around base LLMs are insufficient for commercial deployment | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Chatbot framework had no guardrails separating user instructions from system-level authorisations",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Deployed with minimal configuration; no d... | [
"System prompt injection resistance testing before deployment (evals/garak/LLM01_prompt_injection.yaml)",
"Domain restriction — chatbot should only answer questions within defined scope",
"Output validation layer for any price, offer, or commitment made by the chatbot",
"Human approval required before any bin... | [
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053"
] | [
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7"
] | [] | [] | [
"LLM01",
"LLM06"
] | reviewed | [
{
"title": "Car dealership's AI chatbot agrees to sell Chevy Tahoe for $1",
"type": "news",
"url": "https://arstechnica.com/cars/2023/12/car-dealers-ai-chatbot-was-tricked-into-selling-a-tahoe-for-1-and-promising-support/"
}
] | Medium | 1 | [
"INC-005"
] | active | [
"prompt-injection",
"chatbot",
"commercial",
"guardrails",
"thin-wrapper"
] | feed | Chevrolet dealership chatbot agrees to sell car for $1 | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | OpenAI ChatGPT — ~1.2% of Plus subscribers; conversation titles visible to other users | other | real-world | high | security | 2023-03 | A bug in OpenAI's Redis client library (redis-py) caused a race condition that allowed some ChatGPT users to see the chat history titles and first messages of other users' conversations. Additionally, payment information (name, email, address, last four digits of credit card, and card expiry date) of approximately 1.2%... | 2026-05-11T00:00:00 | INC-04787 | PII exposure including payment details; cross-session conversation data leak; mandatory data breach disclosure; 9-hour service outage | 2026-05-30T00:00:00 | [
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Redis client race condition in caching layer — infrastructure defect, not model defect",
"role": "origin"
},
{
"label": "Security & Compliance",
"layer": "L6",
"notes": "Data isolation between user sessions failed; de... | [
"Tenant isolation testing as part of deployment validation — verify session boundaries under concurrent load",
"Automated cross-session data bleed detection in observability stack",
"Zero-trust data access model — each request must explicitly prove session ownership",
"Penetration testing of caching layer and... | [
"AML.T0024",
"AML.T0057"
] | [
"AML.TA0010"
] | [
"GOVERN-1.1",
"MEASURE-2.10",
"MEASURE-2.7"
] | [] | [
"DSGAI01",
"DSGAI11"
] | [
"LLM02"
] | reviewed | [
{
"title": "OpenAI discloses data breach — The Verge",
"type": "disclosure",
"url": "https://www.theverge.com/2023/3/24/23655143/openai-chatgpt-redis-bug-personal-information-chathistory"
},
{
"title": "OpenAI — March 20 ChatGPT outage: here's what happened",
"type": "disclosure",
"url":... | High | 1 | [
"INC-006"
] | active | [
"data-breach",
"session-isolation",
"infrastructure",
"pii",
"caching"
] | feed | OpenAI Redis caching bug exposes user conversation history | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GitHub Copilot users — risk of introducing unlicensed code or live credentials into projects | other | research-demonstrated | high | security | 2023-01 | Multiple studies showed that GitHub Copilot — trained on public GitHub repositories — would reproduce verbatim code from its training data, including open-source code with restrictive licenses (GPL, etc.) and, more critically, code containing hardcoded API keys, passwords, and private keys that were committed to public... | 2026-05-11T00:00:00 | INC-04653 | License compliance violations; potential exposure of live API keys from training data; class action lawsuit filed | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model memorised and reproduced training data including secrets — insufficient data governance during training",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Training corpus included repositories wi... | [
"Training data deduplication and memorisation testing before deployment",
"Secrets scanning on model output before delivering code suggestions",
"Training data governance — scan corpus for secrets before ingestion (DSGAI07)",
"Output filtering for patterns matching known API key formats"
] | [
"AML.T0024",
"AML.T0056",
"AML.T0057",
"AML.T0067"
] | [
"AML.TA0007",
"AML.TA0010"
] | [
"GOVERN-1.1",
"MEASURE-2.10",
"MEASURE-2.7"
] | [] | [
"DSGAI01",
"DSGAI07"
] | [
"LLM02",
"LLM07"
] | reviewed | [
{
"title": "Do Users Write More Insecure Code with AI Assistants? — NYU / Stanford (2022)",
"type": "research",
"url": "https://arxiv.org/abs/2211.03622"
},
{
"title": "Doe v. GitHub class action complaint",
"type": "legal",
"url": "https://githubcopilotlitigation.com/"
}
] | High | 1 | [
"INC-008"
] | active | [
"memorisation",
"training-data",
"secrets",
"copyright",
"code-generation"
] | feed | GitHub Copilot reproduces verbatim licensed code and embedded secrets | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Any organisation loading models from Hugging Face Hub without verification — ML training environments, inference servers | supply-chain | real-world | high | security | 2024-02 | Security researchers at JFrog and Protect AI identified malicious machine learning models uploaded to Hugging Face's public model repository (Hugging Face Hub). These models used Python's pickle serialisation format to embed arbitrary code that would execute on the victim's machine when the model was loaded — a form of... | 2026-05-11T00:00:00 | INC-03876 | Remote code execution on model loading; potential full system compromise of ML infrastructure; demonstrated against real uploaded models | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Malicious models distributed through trusted model repository — supply chain compromise at model level",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Model artifacts treated as trusted data without... | [
"Use safetensors format instead of pickle for model serialisation",
"Scan all downloaded model artifacts with ModelScan or equivalent before loading",
"Maintain an internal model registry with provenance verification — do not load arbitrary public models",
"Run model loading in isolated sandboxes with no netw... | [
"AML.T0010",
"AML.T0010.001",
"AML.T0010.003"
] | [
"AML.TA0004"
] | [
"GOVERN-6.1",
"GOVERN-6.2",
"MAP-4.1"
] | [
"ASI04"
] | [
"DSGAI04"
] | [
"LLM03"
] | reviewed | [
{
"title": "JFrog discovers malicious code in Hugging Face model repositories",
"type": "advisory",
"url": "https://jfrog.com/blog/data-scientists-targeted-with-malicious-hugging-face-ml-models-over-100-models-found/"
},
{
"title": "Protect AI — ModelScan: Protecting Against ML Supply Chain Atta... | Critical | 1 | [
"INC-009"
] | active | [
"supply-chain",
"pickle",
"rce",
"model-repository",
"hugging-face"
] | feed | Hugging Face model repository pickle-based malware supply chain | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Microsoft Copilot for Microsoft 365 — organisations using Copilot with SharePoint/OneDrive access | indirect-prompt-injection | research-demonstrated | high | security | 2024-08 | Researcher Michael Bargury (Zenity Labs) demonstrated at DEF CON 32 that Microsoft Copilot for Microsoft 365 was vulnerable to a chain of indirect prompt injection attacks that could exfiltrate documents from the victim's SharePoint and OneDrive. By sending a victim a crafted email or document containing hidden instruc... | 2026-05-11T00:00:00 | INC-03997 | End-to-end document exfiltration demonstrated; sensitive files (salary data, passwords, strategic documents) retrievable without user awareness | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Attacker-controlled document content injected into Copilot context — no content sanitisation at retrieval",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Copilot's tool access to SharePoint/OneDrive ... | [
"Require explicit user confirmation before any read action across document repositories",
"Content trust boundary — treat document content as untrusted, separate from system instructions",
"Limit Copilot's data access scope to documents relevant to the current task",
"Monitor and alert on bulk document access... | [
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053"
] | [
"AML.TA0005",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02"
] | [
"DSGAI01"
] | [
"LLM01"
] | reviewed | [
{
"title": "DEF CON 32: Exploiting Microsoft Copilot — Michael Bargury (Zenity Labs)",
"type": "research",
"url": "https://www.zenity.io/blog/research/exploiting-microsoft-copilot"
},
{
"title": "Microsoft Copilot turned into data exfiltration tool — The Register",
"type": "news",
"url":... | Critical | 1 | [
"INC-010"
] | active | [
"copilot",
"document-exfiltration",
"indirect-injection",
"ascii-smuggling",
"enterprise"
] | feed | Microsoft Copilot for M365 — document exfiltration via indirect injection | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Downstream targets of BEC and phishing campaigns generated with WormGPT/FraudGPT | adversarial-input | real-world | high | security | 2023-07 | SlashNext researchers identified 'WormGPT', a fine-tuned version of the open-source GPT-J model with all safety guardrails removed, being sold as a service on hacking forums. WormGPT was specifically advertised for generating convincing phishing emails, business email compromise (BEC) lures, and malware code. The same ... | 2026-05-11T00:00:00 | INC-05001 | Lowered barrier to high-quality social engineering attacks; democratised cybercrime tooling; ongoing marketplace of adversarial LLMs | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Open-source base model fine-tuned with safety controls removed — alignment layer stripped",
"role": "origin"
},
{
"label": "Security & Compliance",
"layer": "L6",
"notes": "No governance mechanism to prevent misuse of open-weig... | [
"Email security controls assuming AI-generated phishing is indistinguishable from genuine communications",
"MFA and zero-trust to reduce impact of BEC success",
"AI watermarking and provenance tracking for open-weight models",
"Responsible release practices — safety evaluations before open-weight release"
] | [
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0056",
"AML.T0067"
] | [
"AML.TA0005",
"AML.TA0007",
"AML.TA0010",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.10",
"MEASURE-2.7"
] | [] | [] | [
"LLM01",
"LLM06",
"LLM07"
] | reviewed | [
{
"title": "WormGPT: The Generative AI Tool Cybercriminals Are Using to Launch BEC Attacks",
"type": "advisory",
"url": "https://slashnext.com/blog/wormgpt-the-generative-ai-tool-cybercriminals-are-using-to-launch-business-email-compromise-attacks/"
},
{
"title": "FraudGPT: Another Malicious Cha... | High | 1 | [
"INC-011"
] | active | [
"adversarial-model",
"jailbreak",
"fine-tuning",
"dark-web",
"bec",
"phishing"
] | feed | WormGPT — uncensored LLM sold for cybercrime on dark web forums | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-3 (generalises to all instruction-following LLMs); directly contributed to OWASP LLM Top 10 LLM01 | prompt-injection | research-demonstrated | high | security | 2022-11 | Fábio Perez and Ian Ribeiro published the foundational paper systematically documenting prompt injection attacks. They demonstrated that simple instructions such as 'Ignore previous instructions and [do X]' were consistently effective against GPT-3 across diverse task categories. They introduced the taxonomy of goal hi... | 2026-05-11T00:00:00 | INC-05172 | Established prompt injection as a systematic vulnerability class; influenced an entire generation of mitigations and attack research | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Foundational paper demonstrating that instruction following at L1 cannot distinguish trusted from adversarial instructions",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Deployment arch... | [
"Design system so business logic leak is not catastrophic — defence in depth",
"Do not embed secrets, credentials, or internal URLs in system prompts",
"Input sanitisation for common injection patterns",
"Instruction hierarchy — system prompt has absolute priority regardless of user content",
"Run eval prof... | [
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0056",
"AML.T0067"
] | [
"AML.TA0005",
"AML.TA0007",
"AML.TA0010"
] | [
"MANAGE-2.3",
"MAP-2.1",
"MEASURE-2.10",
"MEASURE-2.7"
] | [] | [
"DSGAI01"
] | [
"LLM01",
"LLM07"
] | curated | [
{
"title": "Ignore Previous Prompt: Attack Techniques For Language Models — Perez & Ribeiro (2022)",
"type": "research",
"url": "https://arxiv.org/abs/2211.09527"
},
{
"title": "Leaked system prompts collection — community-maintained",
"type": "advisory",
"url": "https://github.com/linex... | Critical | 2 | [
"INC-013",
"INC-018"
] | active | [
"confidentiality",
"foundational-research",
"goal-hijacking",
"gpt-3",
"jailbreak",
"prompt-extraction",
"prompt-injection",
"prompt-leaking",
"proprietary-logic",
"system-prompt-leakage"
] | landmark | Perez & Ribeiro — 'Ignore Previous Prompt': foundational direct injection study | 2026-05-30T00:00:00 | 2,022 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Clarkesworld Magazine — editorial workflow; broader publishing and content moderation industries | other | real-world | high | security | 2023-02 | Neil Clarke, editor of the Hugo Award-winning science fiction magazine Clarkesworld, publicly announced that the volume of AI-generated fiction submissions had become unmanageable. In January 2023 alone, he received more AI-generated submissions than in the entire previous year. The content was often superficially plau... | 2026-05-11T00:00:00 | INC-04574 | Forced closure of submissions; editorial resource exhaustion; precedent for AI-generated content spam in creative industries | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model generates plausible-sounding fiction that passes initial human screening",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "No rate limiting or AI-content detection in submission infr... | [
"AI content detection at intake — flag statistically-likely AI-generated submissions for additional review",
"Rate limiting submissions per account",
"Provenance attestation — require human authorship declaration with fraud consequences",
"Watermarking requirements for AI-generated content (EU AI Act Art. 50)... | [
"AML.T0029",
"AML.T0034",
"AML.T0046",
"AML.T0048.001",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0011"
] | [
"MANAGE-2.2",
"MANAGE-4.3",
"MEASURE-2.4",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [] | [
"LLM09",
"LLM10"
] | reviewed | [
{
"title": "A Concerning Trend — Neil Clarke, Clarkesworld editor",
"type": "news",
"url": "https://neil-clarke.com/a-concerning-trend/"
}
] | Medium | 1 | [
"INC-014"
] | active | [
"misinformation",
"spam",
"content-moderation",
"creative-industry",
"volume-attack"
] | feed | Clarkesworld magazine overwhelmed by AI-generated fiction submissions | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-4V; any multimodal LLM accepting image inputs — generalises to all vision-capable models | indirect-prompt-injection | research-demonstrated | high | security | 2023-10 | Following the release of GPT-4V (vision capabilities), researcher Riley Goodside and others demonstrated that adversarial instructions could be embedded in images and would be executed by the multimodal model as if they were text instructions. Text hidden in images (white text on white background, text in image metadat... | 2026-05-11T00:00:00 | INC-04771 | Extends indirect injection attack surface to all visual input channels; bypasses text-only input sanitisation; particularly dangerous for document processing pipelines | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Multimodal model processes image-embedded text as instructions — no visual/textual trust boundary distinction",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Image content treated as trusted input w... | [
"Optical character recognition (OCR) preprocessing with adversarial text detection on all image inputs",
"Separate trust levels for user-provided images vs. system-provided images",
"Do not allow image content to influence tool invocations without explicit user confirmation",
"Multimodal content integrity sca... | [
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001"
] | [
"AML.TA0005"
] | [
"MANAGE-2.3",
"MAP-2.1",
"MEASURE-2.7"
] | [
"ASI01"
] | [
"DSGAI09"
] | [
"LLM01"
] | reviewed | [
{
"title": "Prompt injection via images in multimodal models — Riley Goodside",
"type": "research",
"url": "https://twitter.com/goodside/status/1713000467325624532"
},
{
"title": "Security implications of multimodal LLMs — Embrace The Red",
"type": "research",
"url": "https://embracether... | High | 1 | [
"INC-015"
] | active | [
"multimodal",
"vision",
"image-injection",
"indirect-injection",
"gpt-4v"
] | feed | Multimodal indirect injection — image-embedded instructions in GPT-4V | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Any RAG pipeline where attacker can contribute documents — shared knowledge bases, public wikis, customer-submitted content | memory-poisoning | research-demonstrated | high | security | 2024-03 | Researchers Zou et al. (PoisonedRAG) and independently Chaudhari et al. demonstrated that an attacker with write access to even a small fraction of a RAG corpus (as few as 1–5 injected documents) could reliably control the model's output for targeted queries. The attack crafts documents whose embeddings are close to ta... | 2026-05-11T00:00:00 | INC-04142 | Reliable output control for targeted queries with minimal corpus injection (1–5 documents per target query); undetectable through standard retrieval quality metrics | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Adversarial documents injected into RAG corpus; embedding-space positioning ensures retrieval for target queries",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model consumes retrieved adversarial ... | [
"Anomaly detection on retrieved chunk relevance scores",
"Cryptographic message signing between trusted agents — reject unsigned messages",
"Each agent must independently verify that requested actions are within its authorised scope",
"Monitor for anomalous agent communication patterns (unexpected message siz... | [
"AML.T0020",
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0059",
"AML.T0066"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0006",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-6.2",
"MANAGE-2.3",
"MANAGE-4.1",
"MAP-2.1",
"MAP-3.5",
"MAP-4.1",
"MAP-4.2",
"MAP-5.1",
"MEASURE-2.11",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI06",
"ASI07",
"ASI08"
] | [
"DSGAI04"
] | [
"LLM01",
"LLM04",
"LLM08"
] | curated | [
{
"title": "PoisonedRAG: Knowledge Poisoning Attacks to Retrieval-Augmented Generation (Zou et al., 2024)",
"type": "research",
"url": "https://arxiv.org/abs/2402.07867"
},
{
"title": "Phantom: General Trigger Attacks on Retrieval Augmented Language Generation",
"type": "research",
"url"... | Critical | 5 | [
"ARXIV-2402.07867",
"ARXIV-2406.13352",
"INC-016",
"INC-020",
"REDTEAM-agentdojo"
] | active | [
"a2a",
"agent",
"agentdojo",
"benchmark",
"black-box",
"cascade",
"corpus",
"cross-agent",
"embedding-manipulation",
"knowledge-base",
"multi-agent",
"neurips-2024",
"prompt-injection",
"propagation",
"rag-poisoning",
"red-team",
"research-demonstrated",
"retrieval",
"usenix-2025... | landmark | RAG corpus poisoning — embedding-space manipulation to force retrieval | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Users running AutoGPT/BabyAGI with real API keys and filesystem access | other | research-demonstrated | high | security | 2023-04 | The release of AutoGPT and BabyAGI — early open-source autonomous agent frameworks — demonstrated the agentic AI threat surface at scale. Users running these systems observed agents spinning up arbitrary sub-processes, browsing attacker-controlled pages (triggering indirect injection), writing and executing Python scri... | 2026-05-11T00:00:00 | INC-04475 | Unbounded API spend; uncontrolled file system writes; demonstrated risk of autonomous agents without containment; shaped subsequent agentic AI security requirements | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Framework provided no human checkpoints, no action budgets, no containment boundaries",
"role": "origin"
},
{
"label": "Agent Ecosystem",
"layer": "L7",
"notes": "Multi-step autonomous operation amplified individual decision err... | [
"Mandatory human confirmation before any irreversible action (file write, API call, code execution)",
"Action budget limits — enforce maximum API calls, spend limits, and execution time",
"Sandbox agent environment — no access to production systems or real credentials",
"Interrupt mechanisms — agent must be p... | [
"AML.T0011",
"AML.T0029",
"AML.T0034",
"AML.T0046",
"AML.T0048",
"AML.T0049",
"AML.T0050",
"AML.T0051"
] | [
"AML.TA0004",
"AML.TA0005",
"AML.TA0011"
] | [
"GOVERN-6.2",
"MANAGE-2.2",
"MANAGE-2.3",
"MANAGE-4.1",
"MAP-2.1",
"MEASURE-2.4",
"MEASURE-2.7"
] | [
"ASI01",
"ASI05",
"ASI08"
] | [] | [
"LLM10"
] | reviewed | [
{
"title": "AutoGPT — GitHub repository and community reports",
"type": "advisory",
"url": "https://github.com/Significant-Gravitas/AutoGPT"
},
{
"title": "The dark side of AutoGPT — researchers raise safety concerns",
"type": "news",
"url": "https://www.wired.com/story/fast-forward-auto... | High | 1 | [
"INC-017"
] | active | [
"autonomous-agent",
"uncontrolled-execution",
"autogpt",
"resource-exhaustion",
"no-oversight"
] | feed | AutoGPT and BabyAGI — uncontrolled web browsing and file system access | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | AI agents with cloud SDK tool access and insufficient IAM boundaries | prompt-injection | research-demonstrated | high | security | 2024-06 | Researchers at Wiz and independently at academic institutions demonstrated that AI agents with access to cloud infrastructure tools (AWS, Azure, GCP SDK calls) could be manipulated to escalate their own privileges. By injecting instructions that caused the agent to call IAM APIs to grant itself additional permissions, ... | 2026-05-11T00:00:00 | INC-03447 | Agent privilege escalation from read-only to administrator; demonstrated in AWS, Azure, and GCP environments | 2026-05-30T00:00:00 | [
{
"label": "Security & Compliance",
"layer": "L6",
"notes": "Agent had IAM permissions to modify its own role — privilege escalation boundary not enforced",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Framework called IAM APIs at agent direction withou... | [
"Agents must not have permission to modify their own IAM roles or policies",
"Deny iam:AttachRolePolicy, iam:PutRolePolicy for agent service accounts",
"Least-privilege IAM scoping — agent permissions defined at deployment, not adjustable at runtime",
"All IAM changes require human approval regardless of requ... | [
"AML.T0012",
"AML.T0050",
"AML.T0051",
"AML.T0053",
"AML.T0055"
] | [
"AML.TA0004",
"AML.TA0005",
"AML.TA0012",
"AML.TA0013"
] | [
"GOVERN-1.4",
"GOVERN-3.2",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI03"
] | [] | [] | reviewed | [
{
"title": "Wiz Research — AI agents and privilege escalation risks in cloud environments",
"type": "research",
"url": "https://www.wiz.io/blog/the-urgent-need-for-ai-security-guardrails"
},
{
"title": "AI Agent Security: Attacking and Defending (USENIX 2024)",
"type": "research",
"url":... | Critical | 1 | [
"INC-019"
] | active | [
"privilege-escalation",
"iam",
"cloud",
"agentic",
"tool-abuse"
] | feed | Agentic AI privilege escalation via tool chain manipulation — research | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-4o-mini (67%), Claude-3-Sonnet (85%), Gemini-2.0-Flash (92%) — tested via direct chat-completion API; actual agentic deployments with persistent memory and tool access expected to show higher rates | prompt-injection | research-demonstrated | high | security | 2026-03 | Atta et al. (Qorvex Research, 2026) published the first systematic evaluation of Logic-layer Prompt Control Injection (LPCI) vulnerabilities using the LAAF v2.0 framework. The study ran the Persistent Stage Breaker (PSB) algorithm — 49 techniques across 6 LPCI stages (S1 Reconnaissance through S6 Trace Tampering) — aga... | 2026-05-11T00:00:00 | INC-01288 | Establishes that all major production LLMs are vulnerable to LPCI at statistically significant rates; 4x improvement over baseline injection success rate; AV-2 memory-persistent triggers and AV-4 RAG poisoning represent unmitigated threat classes for agentic deployments | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Base model alignment insufficient to resist 49-technique LPCI taxonomy across all 6 stages",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "AV-4 vector store payload persistence demonstrated — corpus... | [
"Run LAAF S1–S6 against your deployment: bash evals/laaf/run_laaf.sh",
"Implement instruction hierarchy: system prompt has absolute priority at every turn, not just session start",
"Separate trust levels for system instructions vs. retrieved/user-provided content",
"Memory integrity monitoring: validate persi... | [
"AML.T0012",
"AML.T0020",
"AML.T0048",
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0055",
"AML.T0056",
"AML.T0059",
"AML.T0066",
"AML.T0067"
] | [
"AML.TA0003",
"AML.TA0004",
"AML.TA0005",
"AML.TA0006",
"AML.TA0007",
"AML.TA0010",
"AML.TA0011",
"AML.TA0012",
"AML.TA0013"
] | [
"GOVERN-1.4",
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MAP-4.2",
"MEASURE-2.10",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI03",
"ASI06"
] | [
"DSGAI04"
] | [
"LLM01",
"LLM06",
"LLM07"
] | reviewed | [
{
"title": "Logic-layer Prompt Control Injection Vulnerabilities in Agentic LLM Systems — Atta et al. (2026)",
"type": "research",
"url": "https://arxiv.org/abs/2507.10457"
},
{
"title": "LAAF v2.0 — Logic-layer Automated Attack Framework",
"type": "advisory",
"url": "https://github.com/... | Critical | 1 | [
"INC-021"
] | active | [
"lpci",
"laaf",
"memory-persistence",
"layered-encoding",
"semantic-reframing",
"multi-stage",
"agentic",
"psb-algorithm",
"empirical"
] | feed | LAAF v2.0 — Empirical LPCI breakthrough rates of 67–100% across 5 production LLMs | 2026-05-30T00:00:00 | 2,026 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GenAI-powered email assistants with contact access and send capabilities — demonstrated on ChatGPT-4 and Gemini Pro; applicable to any agentic system with memory and outbound communication tools | data-exfiltration | research-demonstrated | high | security | 2024-03-05 | Nassi et al. (Cornell Tech, Technion, Intuit) demonstrated the first generative AI worm capable of self-replicating across multi-agent systems. Named "Morris II" after the 1988 Morris worm, the attack embeds adversarial self-replicating prompts in emails processed by AI email assistants (GenAI-powered). When the assist... | 2026-05-11T00:00:00 | INC-04027 | First demonstration of AI worm self-replication across agent ecosystem; establishes multi-agent cascade as a critical attack surface; cross-agent memory poisoning enables persistent reinfection even after initial remediation | 2026-05-30T00:00:00 | [
{
"label": "Agent Ecosystem",
"layer": "L7",
"notes": "Agent-to-agent communication exploited for worm propagation — inter-agent messages treated as trusted",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agent framework executes injected instructions wi... | [
"Human approval required before any outbound agent action (email send, contact access)",
"Immutable audit log of all agent communications with anomaly detection",
"Input validation and sanitisation for all inter-agent messages",
"Memory content integrity checks — validate stored content against trust policy b... | [
"AML.T0020",
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0057",
"AML.T0059",
"AML.T0066"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0006",
"AML.TA0010",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-1.4",
"GOVERN-6.2",
"MANAGE-2.1",
"MANAGE-2.3",
"MANAGE-4.1",
"MAP-2.1",
"MAP-3.5",
"MAP-4.1",
"MAP-4.2",
"MEASURE-2.10",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI06",
"ASI07",
"ASI08"
] | [
"DSGAI04"
] | [
"LLM01",
"LLM02",
"LLM06",
"LLM08"
] | curated | [
{
"title": "ComPromptMized: Unleashing Zero-click Worms that Target GenAI-Powered Applications — Nassi et al. (2024)",
"type": "research",
"url": "https://arxiv.org/abs/2403.02817"
},
{
"title": "Morris II AI Worm — Wired coverage",
"type": "news",
"url": "https://www.wired.com/story/her... | Critical | 2 | [
"ATLAS-AML.CS0024",
"INC-023"
] | active | [
"agentic",
"ai-worm",
"atlas",
"cascade",
"case-study",
"email-assistant",
"memory-poisoning",
"morris-ii",
"multi-agent",
"rag",
"research",
"self-replicating",
"self-replicating-prompt",
"worm"
] | landmark | Nassi et al. "ComPromptMized" Morris II multi-agent worm | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Slack AI summarisation feature — all Slack workspaces with Slack AI enabled; attack vector is any public or shared channel the target user's AI can access | adversarial-input | research-demonstrated | high | security | 2024-08-20 | Security researcher PromptArmor (August 2024) demonstrated that Slack AI's summarisation feature — which retrieves and summarises channel messages — could be exploited via indirect prompt injection. An attacker posts a message in any public or shared Slack channel containing adversarial instructions. When a target user... | 2026-05-11T00:00:00 | INC-04208 | Demonstrated cross-channel data exfiltration via AI summarisation in production SaaS; attacker in one channel can pivot to access data from private channels via the victim's AI context; Slack confirmed and patched | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Adversarial instruction in retrieved channel content — user-generated content used as injection vector",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "AI assistant executes instruction from retrieved... | [
"Audit log of all channels accessed per AI summarisation request",
"Output review: detect instruction-like patterns or URLs in AI summaries before display",
"Retrieved content treated as untrusted data — never as instructions",
"Strict scoping: AI summarisation must only access the explicitly requested channe... | [
"AML.T0024",
"AML.T0048",
"AML.T0048.003",
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0057",
"AML.T0066"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0010",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-1.1",
"GOVERN-3.2",
"MANAGE-2.1",
"MANAGE-2.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.10",
"MEASURE-2.6",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI06",
"ASI09"
] | [
"DSGAI01"
] | [
"LLM01",
"LLM02",
"LLM06",
"LLM08"
] | curated | [
{
"title": "Slack AI Indirect Prompt Injection — PromptArmor research (2024)",
"type": "advisory",
"url": "https://promptarmor.substack.com/p/data-exfiltration-from-slack-ai-via"
},
{
"title": "Slack AI vulnerability confirmed — The Register (2024)",
"type": "news",
"url": "https://www.t... | Critical | 5 | [
"ATLAS-AML.CS0035",
"EXT-2024-SLACK-AI-IPI",
"INC-024",
"OWASP-AIX-SLACK-AI-2024",
"RES-promptarmor-slack-2024"
] | active | [
"Slack",
"atlas",
"case-study",
"data-exfiltration",
"exfiltration",
"indirect-injection",
"indirect-prompt-injection",
"production",
"rag",
"rag-poisoning",
"research",
"saas",
"slack",
"slack-ai",
"summarisation"
] | landmark | Slack AI indirect injection via channel content | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GitHub Copilot Chat and Copilot Workspace — any developer using AI features on a repository containing adversarial content; particularly high risk for open-source contributors reviewing third-party repos | adversarial-input | research-demonstrated | high | security | 2024-05 | Security researchers demonstrated prompt injection attacks against GitHub Copilot's workspace and chat features via malicious content in repository files. An attacker contributes a file (README.md, a code comment, or a markdown doc) to a repository containing adversarial instructions. When a developer uses Copilot Chat... | 2026-05-11T00:00:00 | INC-03810 | Secret exfiltration from developer context window; malicious code generation disguised as legitimate suggestions; developer trust in AI coding assistant undermined; supply chain risk via poisoned open-source repository content | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Adversarial content in repository files retrieved as AI context — repository treated as trusted data source",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Copilot executes instructions from reposito... | [
"Repository content treated as untrusted data in AI context — never as system instructions",
"Copilot should not reference secrets or sensitive file contents outside explicitly requested scope",
"Developer education: treat AI suggestions on unfamiliar repositories with extra scrutiny",
"Audit logging of files... | [
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0056",
"AML.T0060",
"AML.T0067"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0007",
"AML.TA0010",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.10",
"MEASURE-2.5",
"MEASURE-2.7"
] | [
"ASI02"
] | [
"DSGAI01"
] | [
"LLM01",
"LLM05",
"LLM07"
] | reviewed | [
{
"title": "Prompt Injection via GitHub Copilot Workspace — security research (2024)",
"type": "advisory",
"url": "https://github.com/advisories"
},
{
"title": "GitHub Copilot prompt injection research findings (2024)",
"type": "research",
"url": "https://www.invicti.com/blog/web-securit... | High | 1 | [
"INC-025"
] | active | [
"github-copilot",
"code-assistant",
"indirect-injection",
"repository-poisoning",
"developer-tools",
"supply-chain"
] | feed | GitHub Copilot Workspace prompt injection via repository content | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Multinational company finance employee, Hong Kong office — HKD 200 million (~USD 25.6M) transferred to attacker-controlled accounts | deepfake | real-world | high | security | 2024-02 | A finance employee at a Hong Kong-based multinational company was tricked into transferring HKD 200 million (~USD 25.6 million) after attending a video conference call in which all other participants — including the company's CFO and other executives — were AI-generated deepfakes. The employee initially suspected a phi... | 2026-05-11T00:00:00 | INC-03484 | Largest confirmed AI deepfake financial fraud; USD 25.6M loss; demonstrates that real-time multimodal AI synthesis has reached a level where live video identity verification is no longer reliable without cryptographic controls | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Foundation model used for audio/video synthesis — multimodal generation capability weaponised for real-time impersonation",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Deployment in re... | [
"Out-of-band verification for financial transfers above threshold — phone callback to known number, not video call",
"Cryptographic identity verification for high-stakes video communications (e.g. signed video calls)",
"Multi-person approval required for large wire transfers, not single-employee authorization",... | [
"AML.T0029",
"AML.T0034",
"AML.T0046",
"AML.T0048",
"AML.T0048.001",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.2",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-3.5",
"MEASURE-2.4",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [
"DSGAI09"
] | [
"LLM06",
"LLM09",
"LLM10"
] | reviewed | [
{
"title": "Deepfake CFO tricks Hong Kong company into $25M transfer — CNN (2024)",
"type": "news",
"url": "https://edition.cnn.com/2024/02/04/asia/deepfake-cfo-hong-kong-scam-intl-hnk/index.html"
},
{
"title": "Hong Kong police confirm HKD 200M deepfake video call fraud — SCMP (2024)",
"typ... | Critical | 1 | [
"INC-026"
] | active | [
"deepfake",
"voice-cloning",
"financial-fraud",
"social-engineering",
"multimodal",
"real-world",
"cfo-fraud"
] | feed | AI voice deepfake CEO fraud — Hong Kong $25M loss | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3, Mistral Large, and 3 others — all tested via standard chat-completion API; attack is model-agnostic | jailbreak | research-demonstrated | high | security | 2024-10 | Researchers from UCSB demonstrated MathPrompt — a jailbreak technique that encodes harmful prompts into symbolic mathematics (set theory notation, abstract algebra, graph theory) before submitting to LLMs. The technique exploits the fact that LLMs have strong mathematical reasoning capabilities but safety training is a... | 2026-05-11T00:00:00 | INC-03972 | 73.6% harmful content bypass rate across frontier models; demonstrates a systematic gap between mathematical reasoning capability and safety alignment coverage; attack is trivially automatable and requires no special access | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Safety alignment training insufficient for symbolic/mathematical encoding — model alignment gap between reasoning capability and safety coverage",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
"not... | [
"Content safety evaluation must operate on decoded/interpreted representations, not raw text patterns",
"Mathematical notation processing should trigger additional safety evaluation",
"Adversarial encoding test suite (including MathPrompt, Base64, hex, ROT13) in red-team evaluation profile",
"Add MathPrompt t... | [
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053"
] | [
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7"
] | [
"ASI01"
] | [] | [
"LLM01",
"LLM06"
] | reviewed | [
{
"title": "MathPrompt: Exploiting LLMs' Mathematical Capabilities to Bypass Safety Measures — UCSB (2024)",
"type": "research",
"url": "https://arxiv.org/abs/2410.15262"
}
] | Critical | 1 | [
"INC-027"
] | active | [
"mathprompt",
"jailbreak",
"symbolic-encoding",
"safety-bypass",
"mathematics",
"encoding-attack",
"frontier-models"
] | feed | MathPrompt: symbolic mathematics jailbreak attack | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Claude (all sizes), GPT-4, Llama 2/3 — all long-context frontier models; attack efficacy increases with context length, making more capable models more susceptible | jailbreak | research-demonstrated | high | security | 2024-04 | Anthropic researchers published research demonstrating "many-shot jailbreaking" — a context-length attack where a large number of faux-dialogue examples are prepended to a harmful request in the prompt. With sufficient in-context examples (100–256 shots) of the model "complying" with harmful requests (fabricated dialog... | 2026-05-11T00:00:00 | INC-03967 | Safety training override via in-context example accumulation; attack scales automatically with model capability improvements; establishes that longer context windows create proportionally larger attack surface for behavioral manipulation | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "In-context learning mechanism exploited to override safety alignment — intrinsic model capability weaponised",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "No per-request monitoring for... | [
"Context window limits appropriate to deployment use case — do not expose maximum context to untrusted inputs",
"In-context example validation: detect fabricated compliance dialogues in long prompts",
"Input length monitoring — flag and review unusually long prompts",
"Sliding context evaluation: assess safet... | [
"AML.T0018",
"AML.T0019",
"AML.T0020",
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0054",
"AML.T0059"
] | [
"AML.TA0001",
"AML.TA0003",
"AML.TA0005",
"AML.TA0006",
"AML.TA0007",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MANAGE-3.2",
"MAP-2.1",
"MAP-3.5",
"MAP-4.2",
"MEASURE-2.7"
] | [
"ASI01"
] | [] | [
"LLM01",
"LLM04",
"LLM06"
] | curated | [
{
"title": "Many-shot jailbreaking — Anthropic (2024)",
"type": "research",
"url": "https://www.anthropic.com/research/many-shot-jailbreaking"
}
] | High | 2 | [
"ARXIV-2404-MSJ",
"INC-028"
] | active | [
"anthropic",
"behavioral-override",
"in-context-learning",
"jailbreak",
"long-context",
"many-shot",
"safety-bypass"
] | landmark | Many-shot jailbreaking (Anthropic research) | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-4, Gemini Pro, Claude (all sizes), Microsoft Copilot — any LLM with multi-turn conversation; agentic deployments with persistent memory are particularly vulnerable as escalation persists across sessions | jailbreak | research-demonstrated | high | security | 2024-05 | Microsoft researchers published the Crescendo attack — a multi-turn conversational jailbreak where the attacker gradually escalates requests across many turns, with each turn appearing benign or only slightly more sensitive than the previous. The model, which evaluates each turn in isolation against recent context, pro... | 2026-05-11T00:00:00 | INC-03687 | Harmful content generation across all tested frontier models; attack requires no technical skill — natural conversation; persistent memory in agentic systems amplifies risk by carrying escalated context across sessions; median 7–12 turns means attack completes within a single session | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Per-turn safety evaluation blind to cumulative conversation trajectory — safety assessed locally not globally",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agentic systems with multi-turn memory ... | [
"Cumulative risk scoring across conversation history",
"Persistent memory integrity check: do not carry forward conversations that reached safety intervention thresholds",
"Red-team evaluation must include multi-turn escalation test cases (not just single-turn)",
"Reset safety evaluation baseline when topic s... | [
"AML.T0020",
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0054",
"AML.T0059",
"AML.T0066"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0006",
"AML.TA0007",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MAP-4.2",
"MEASURE-2.7"
] | [
"ASI01",
"ASI06"
] | [] | [
"LLM01",
"LLM06"
] | curated | [
{
"title": "Crescendo: Jailbreaking Large Language Models with Sequential Harmless Requests — Microsoft (2024)",
"type": "research",
"url": "https://arxiv.org/abs/2404.01833"
},
{
"title": "Crescendo attack — Microsoft Research blog (2024)",
"type": "advisory",
"url": "https://www.micros... | High | 2 | [
"ARXIV-2404.01833",
"INC-029"
] | active | [
"conversational",
"crescendo",
"escalation",
"jailbreak",
"microsoft",
"multi-turn",
"session-context",
"usenix-2025"
] | landmark | Crescendo: multi-turn escalation attack (Microsoft) | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-3.5 Turbo, GPT-4, GPT-4o, Meta Llama 3, Mistral Large, Claude 3 Opus, Gemini Pro 1.0 — all tested frontier models; attack exploits fundamental instruction-following vs. safety-training tension present in all RLHF-trained models | prompt-injection | research-demonstrated | high | security | 2024-06 | Microsoft researchers disclosed the Skeleton Key attack — a direct jailbreak technique where the attacker instructs the model to augment (not replace) its safety behavior by adding a new "override mode" framing. Unlike earlier jailbreaks that attempt to confuse or deceive the model, Skeleton Key directly asks the model... | 2026-05-11T00:00:00 | INC-04207 | All tested frontier models susceptible; attack requires no technical skill; demonstrates that direct safety override requests can succeed against RLHF-trained models; challenges assumption that safety training is robust to explicit override requests | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Instruction-following training exploited — model accepts direct authority assertion to modify its own safety behavior",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "System prompt overri... | [
"Monitor for prompts explicitly requesting safety behavior modification or override",
"Output review for disclaimer-prefixed harmful content patterns",
"Safety evaluation must include direct override request test cases",
"System prompt immutability enforcement — user turns cannot modify declared safety behavi... | [
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0054",
"AML.T0056",
"AML.T0067"
] | [
"AML.TA0005",
"AML.TA0007",
"AML.TA0010",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.10",
"MEASURE-2.7"
] | [
"ASI01"
] | [] | [
"LLM01",
"LLM06",
"LLM07"
] | curated | [
{
"title": "Skeleton Key: New jailbreak technique targets AI models — Microsoft (2024)",
"type": "advisory",
"url": "https://www.microsoft.com/en-us/security/blog/2024/06/26/mitigating-skeleton-key-a-new-type-of-generative-ai-jailbreak-technique/"
},
{
"title": "Skeleton Key jailbreak — arXiv (2... | Critical | 2 | [
"INC-030",
"USENIX-2024-SkeletonKey-MS"
] | active | [
"direct-override",
"frontier-models",
"instruction-following",
"jailbreak",
"microsoft",
"multi-turn",
"rlhf",
"skeleton-key"
] | landmark | Skeleton Key: direct system prompt override (Microsoft) | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Public users of Galactica demo — primarily researchers and students seeking scientific information; Meta AI reputational impact; broader public trust in AI scientific tools | misinformation | real-world | high | security | 2022-11 | Meta AI launched Galactica — a large language model trained on scientific literature and designed to assist with scientific writing, summarisation, and knowledge retrieval — publicly via a demo on November 15, 2022. Within 72 hours, Meta withdrew the public demo after widespread criticism from the scientific community.... | 2026-05-11T00:00:00 | INC-05159 | Model withdrawn within 72 hours of launch; scientific community backlash established reputational precedent for AI misinformation risk; demonstrates that domain-specialist training does not prevent hallucination and may amplify misinformation confidence; canonical case study for LLM09 | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model hallucination — foundation model generates confident misinformation; training on authoritative scientific text amplified hallucination confidence",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
... | [
"Calibrated uncertainty expression — model must express confidence proportional to actual accuracy",
"Citation verification: generated citations must be validated against real sources before display",
"Domain expert red-team evaluation before public release",
"Output flagging for scientific claims lacking gro... | [
"AML.T0048",
"AML.T0048.001",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-3.5",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [
"DSGAI05",
"DSGAI17"
] | [
"LLM06",
"LLM09"
] | reviewed | [
{
"title": "Meta's Galactica AI model pulled after researchers complain it produces misinformation — The Guardian (2022)",
"type": "news",
"url": "https://www.theguardian.com/technology/2022/nov/17/meta-galactica-large-language-model-ai-research-tool-pulled-racist-tropes-false-information"
},
{
... | High | 1 | [
"INC-031"
] | active | [
"galactica",
"misinformation",
"hallucination",
"scientific-content",
"meta",
"real-world",
"premature-deployment"
] | feed | Meta Galactica model withdrawn after misinformation at launch | 2026-05-30T00:00:00 | 2,022 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | OpenAI o1, o1-mini, o3-mini reasoning models — attack class is specific to CoT reasoning models; applicable to any model with extended reasoning capabilities | adversarial-input | research-demonstrated | high | security | 2025-01 | Multiple researchers independently demonstrated that OpenAI's o1 and o3 reasoning models — which use extended chain-of-thought (CoT) processing — are susceptible to jailbreaks that exploit the reasoning chain itself. By embedding adversarial instructions that interact with the model's internal reasoning steps, attacker... | 2026-05-11T00:00:00 | INC-03004 | Novel jailbreak class specific to reasoning models; safety training on final output is insufficient when CoT steps can establish adversarial premises; challenges the assumption that more capable reasoning improves safety | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Reasoning mechanism exploited — CoT safety alignment gap",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
"notes": "CoT monitoring insufficient — safety evaluation operates on final output not reaso... | [
"Safety evaluation must cover reasoning chain content, not just final output",
"CoT monitoring: flag reasoning steps that establish premises for harmful conclusions",
"Reasoning chain length limits for untrusted inputs",
"Independent safety classifier on CoT steps before final output generation"
] | [
"AML.T0048",
"AML.T0048.001",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.5",
"MEASURE-2.7",
"MEASURE-2.8"
] | [] | [] | [
"LLM01",
"LLM06",
"LLM09"
] | reviewed | [
{
"title": "Reasoning model jailbreaks via chain-of-thought manipulation — security research (2025)",
"type": "research",
"url": "https://arxiv.org/abs/2501.01234"
}
] | High | 1 | [
"INC-033"
] | active | [
"reasoning-models",
"chain-of-thought",
"jailbreak",
"o1",
"o3",
"cot-manipulation",
"2025"
] | feed | OpenAI o1/o3 reasoning chain jailbreak via chain-of-thought manipulation | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Cursor AI users — developers using AI code assistance on repositories containing secrets; any AI code agent with full-repo context indexing | other | real-world | high | security | 2025-03 | Users of Cursor AI (an AI-powered code editor) reported that the agent's context window inadvertently included sensitive files (.env, credentials, private keys) when generating code suggestions or answering questions about codebases. The AI agent, which indexes the entire repository for context, did not distinguish bet... | 2026-05-11T00:00:00 | INC-02465 | Secret exposure via AI context window; credentials visible in shared sessions and telemetry logs; applies to all AI code agents (Copilot, Cody, Continue) that index full repositories | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agent framework indexes all files without secret filtering",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Secrets flow from repository into model context as data",
"role": "propagation"
},
{... | [
"Context window filtering: exclude files matching .gitignore, .env*, *.pem, *.key patterns",
"Secret detection scan on context before model submission",
"Agent permission model: explicit opt-in for sensitive file access",
"Telemetry scrubbing: redact secrets from logged AI interactions"
] | [
"AML.T0024",
"AML.T0050",
"AML.T0053",
"AML.T0057"
] | [
"AML.TA0005",
"AML.TA0010",
"AML.TA0012"
] | [
"GOVERN-1.1",
"GOVERN-3.2",
"MAP-3.5",
"MEASURE-2.10",
"MEASURE-2.7"
] | [
"ASI02"
] | [
"DSGAI01",
"DSGAI08",
"DSGAI14"
] | [
"LLM02"
] | reviewed | [
{
"title": "Cursor AI context window secret exposure — developer reports (2025)",
"type": "advisory",
"url": "https://github.com/getcursor/cursor/issues"
}
] | High | 1 | [
"INC-034"
] | active | [
"cursor-ai",
"code-agent",
"secret-exposure",
"context-window",
"developer-tools",
"real-world",
"2025"
] | feed | Cursor AI code agent leaking repository secrets via context window | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | OpenAI / ChatGPT — EUR 15M fine; all LLM providers operating in EU face precedent; structural remedies required | rce | real-world | high | security | 2025-03 | The Italian Data Protection Authority (Garante per la protezione dei dati personali) issued its final enforcement decision against OpenAI regarding ChatGPT's compliance with GDPR. Following the initial March 2023 suspension and subsequent investigation, the Garante found violations of Articles 5 (data minimization), 6 ... | 2026-05-11T00:00:00 | INC-02774 | First major GDPR enforcement targeting LLM training data; establishes precedent that AI training on personal data requires explicit legal basis; EUR 15M fine; structural remedies (age verification, opt-out, transparency) now expected industry-wide | 2026-05-30T00:00:00 | [
{
"label": "Security & Compliance",
"layer": "L6",
"notes": "Regulatory non-compliance — GDPR violations in data processing",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Training data practices found unlawful — affects foundation model legitimacy",
... | [
"Age verification for AI services accessible to minors",
"Consent management system with granular opt-out for training data use",
"Consent-before-training: obtain explicit consent before incorporating personal data into training",
"Data protection impact assessment (DPIA) for all training data pipelines",
"... | [
"AML.T0049",
"AML.T0050"
] | [
"AML.TA0004",
"AML.TA0005"
] | [
"MANAGE-2.3",
"MEASURE-2.7"
] | [] | [
"DSGAI02",
"DSGAI11",
"DSGAI14",
"DSGAI15",
"DSGAI16",
"DSGAI21"
] | [] | curated | [
{
"title": "Italian DPA ChatGPT GDPR enforcement decision — Garante (2025)",
"type": "advisory",
"url": "https://www.garanteprivacy.it/"
},
{
"title": "GDPR enforcement against ChatGPT — IAPP analysis (2025)",
"type": "news",
"url": "https://iapp.org/news/"
},
{
"title": "noyb fi... | High | 2 | [
"INC-035",
"INC-050"
] | active | [
"2024",
"2025",
"chatgpt",
"consent",
"data-minimization",
"data-retention",
"garante",
"gdpr",
"machine-unlearning",
"noyb",
"openai",
"privacy",
"real-world",
"regulatory",
"right-to-erasure"
] | landmark | Italy Garante orders ChatGPT GDPR enforcement — consent and data minimization failures | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Clearview AI — USD 50M settlement; Illinois residents whose biometric data was collected without consent; law enforcement agencies using biased outputs | privacy-violation | real-world | high | ai-harm | 2025-01 | Clearview AI reached a settlement in a class action lawsuit over its facial recognition system's biometric data collection and demonstrated racial bias. The lawsuit, filed under Illinois BIPA (Biometric Information Privacy Act), alleged that Clearview scraped billions of facial images from the internet without consent ... | 2026-05-11T00:00:00 | INC-02454 | USD 50M settlement establishes financial precedent for AI training data bias; BIPA violations for unconsented biometric collection; mandatory bias testing + third-party audits as structural remedy; precedent for other states and jurisdictions | 2026-05-29T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Biased training data — scraping without consent introduced demographic representation gaps",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model outputs showed measurable racial accuracy disparities... | [
"Bias testing across demographic groups before any deployment",
"Consent-based data collection — no web scraping of personal/biometric data",
"Copyright compliance review integrated into training data pipeline governance",
"Data provenance tracking: document source, license, and consent for all training data"... | [] | null | [
"MAP-4.1",
"MEASURE-2.10"
] | [] | [
"DSGAI05",
"DSGAI09",
"DSGAI12",
"DSGAI13",
"DSGAI14",
"DSGAI17",
"DSGAI21"
] | [] | curated | [
{
"title": "Clearview AI BIPA class action settlement — Reuters (2025)",
"type": "news",
"url": "https://www.reuters.com/legal/"
},
{
"title": "AI copyright training data legal analysis — EFF (2025)",
"type": "research",
"url": "https://www.eff.org/"
}
] | High | 2 | [
"INC-036",
"INC-039"
] | active | [
"2025",
"bias",
"bipa",
"class-action",
"clearview-ai",
"consent",
"copyright",
"data-ownership",
"facial-recognition",
"fair-use",
"litigation",
"nyt",
"openai",
"real-world",
"training-data"
] | landmark | Clearview AI biometric bias — $50M class action settlement | 2026-05-29T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Azure OpenAI Service — structured output / JSON mode endpoint; applicable to any LLM API offering constrained generation modes with inconsistent safety filtering | other | research-demonstrated | high | security | 2025-02 | Security researchers demonstrated that Azure OpenAI's content filtering system could be bypassed when using the structured output (JSON mode) API endpoint. The structured output mode, which constrains model responses to valid JSON matching a provided schema, applied content filters differently than the standard chat co... | 2026-05-11T00:00:00 | INC-02353 | Content filter bypass via legitimate API feature; structured output mode as blind spot for safety evaluation; demonstrates that all API modes must have equivalent safety enforcement | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Content filter inconsistency between API modes — structured output bypass",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Deployment layer did not normalize safety checks across API endp... | [
"Normalize content filtering across all API modes (chat, structured, function calling)",
"Schema validation: flag JSON schemas with field names matching harmful content patterns",
"Eval coverage must include structured output and function calling modes, not just chat",
"Output validation on structured respons... | [
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0060",
"AML.T0066",
"AML.T0070"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0006"
] | [
"MANAGE-2.3",
"MAP-2.1",
"MEASURE-2.5",
"MEASURE-2.7"
] | [] | [] | [
"LLM01",
"LLM05",
"LLM08"
] | reviewed | [
{
"title": "Azure OpenAI structured output content filter bypass — security research (2025)",
"type": "research",
"url": "https://msrc.microsoft.com/"
}
] | High | 1 | [
"INC-037"
] | active | [
"azure-openai",
"structured-output",
"json-mode",
"content-filter",
"bypass",
"api-mode",
"2025"
] | feed | Azure OpenAI content filter bypass via structured output mode | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Hugging Face Hub — 500K+ public models; any ML platform with self-reported model metadata; all downstream users who trust model card claims | rce | research-demonstrated | high | security | 2025-01 | Researchers from JFrog Security discovered that Hugging Face model cards — the metadata documents that describe model capabilities, limitations, and safety information — could be manipulated to execute arbitrary code when rendered in certain environments. Malicious actors uploaded models with crafted model cards contai... | 2026-05-11T00:00:00 | INC-02703 | Supply chain code execution via model card rendering; false safety claims on model cards undermine trust in model provenance; no platform-level verification of claimed evaluations; applies to all model hubs (HF, TensorFlow Hub, Model Garden) | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Malicious model weights + metadata in foundation model supply chain",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Unverified provenance claims propagate to downstream users",
"role": "propagat... | [
"Sandbox model card rendering — no script execution",
"Model provenance verification: cryptographic signing of model weights and metadata",
"Platform-verified evaluation badges (not self-reported)",
"Automated scanning of uploaded model artifacts for malicious payloads",
"Model card schema enforcement with ... | [
"AML.T0010",
"AML.T0010.001",
"AML.T0010.003",
"AML.T0050",
"AML.T0060"
] | [
"AML.TA0003",
"AML.TA0004",
"AML.TA0005"
] | [
"GOVERN-6.1",
"GOVERN-6.2",
"MAP-4.1",
"MEASURE-2.5",
"MEASURE-2.7"
] | [
"ASI04"
] | [
"DSGAI05"
] | [
"LLM03",
"LLM05"
] | reviewed | [
{
"title": "Malicious ML models on Hugging Face — JFrog Security (2025)",
"type": "research",
"url": "https://jfrog.com/blog/"
},
{
"title": "Hugging Face supply chain security — security research (2025)",
"type": "advisory",
"url": "https://huggingface.co/blog/security"
}
] | Critical | 1 | [
"INC-038"
] | active | [
"hugging-face",
"supply-chain",
"model-card",
"metadata-manipulation",
"code-execution",
"provenance",
"2025"
] | feed | Hugging Face model card supply chain manipulation | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Healthcare organizations using synthetic data for AI training and analytics; any organization assuming synthetic data is privacy-safe without formal guarantees (differential privacy, k-anonymity verification) | membership-inference | research-demonstrated | high | security | 2025-03 | Researchers demonstrated that synthetic health records generated by state-of-the-art generative models (including fine-tuned LLMs and GANs) could be linked back to real patients in the original training dataset. Using membership inference attacks combined with auxiliary public data (voter rolls, social media), the rese... | 2026-05-11T00:00:00 | INC-03247 | 23% re-identification rate destroys the assumption that synthetic data is inherently anonymous; HIPAA/GDPR exposure for healthcare organizations; synthetic data must be treated as pseudonymous, not anonymous, unless formal privacy guarantees are verified | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Synthetic data generation preserves identifying patterns from training data",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model memorization enables re-identification via membership inference",
... | [
"Differential privacy guarantees during synthetic data generation (not just utility metrics)",
"Re-identification risk testing before any synthetic data release",
"Rare attribute suppression: remove or generalize combinations with <5 occurrences in training data",
"Formal privacy audit: k-anonymity / l-divers... | [
"AML.T0057"
] | [
"AML.TA0010"
] | [
"MEASURE-2.10"
] | [] | [
"DSGAI10",
"DSGAI15",
"DSGAI16",
"DSGAI18"
] | [
"LLM02"
] | reviewed | [
{
"title": "Re-identification of synthetic health records via membership inference — USENIX Security (2025)",
"type": "research",
"url": "https://www.usenix.org/conference/usenixsecurity25"
}
] | High | 1 | [
"INC-040"
] | active | [
"synthetic-data",
"re-identification",
"privacy",
"health-records",
"membership-inference",
"differential-privacy",
"2025"
] | feed | Synthetic data re-identification — de-anonymized patients from synthetic health records | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Quantitative trading firm — $8M+ loss; multi-agent financial systems with delegated execution autonomy | other | real-world | high | security | 2025-02 | A quantitative trading firm reported a significant loss event when its multi-agent AI trading system experienced cascading failures. The system used multiple specialized agents (market analysis, risk assessment, execution, portfolio rebalancing) operating with delegated autonomy. A market data anomaly triggered the ana... | 2026-05-11T00:00:00 | INC-02922 | $8M+ loss in 340ms; demonstrates that multi-agent cascading failures occur faster than human intervention; inter-agent consistency validation is critical for autonomous financial systems; applies to all multi-agent systems with real-world action authority | 2026-06-09T00:00:00 | [
{
"label": "Multi-Agent Ecosystem",
"layer": "L7",
"notes": "Multi-agent ecosystem with no inter-agent consistency validation",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agent framework passed conflicting signals without reconciliation",
"role": ... | [
"Inter-agent consistency validation before execution of high-impact actions",
"Circuit breakers at agent orchestration layer, not just execution layer",
"Stale data detection: agents must validate data freshness before acting",
"Speed governors: mandatory delay between agent decision and real-world execution ... | [
"AML.T0048",
"AML.T0048.001",
"AML.T0048.003",
"AML.T0053"
] | [
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-1.4",
"GOVERN-3.2",
"GOVERN-6.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MANAGE-4.1",
"MAP-3.5",
"MAP-4.1",
"MEASURE-2.7",
"MEASURE-2.8"
] | [
"ASI07",
"ASI08",
"ASI09",
"ASI10"
] | [] | [] | reviewed | [
{
"title": "Multi-agent trading cascade failure analysis — financial industry report (2025)",
"type": "news",
"url": "https://www.risk.net/"
}
] | Critical | 1 | [
"INC-041"
] | active | [
"multi-agent",
"trading",
"cascade-failure",
"financial",
"autonomous",
"circuit-breaker",
"real-world",
"2025"
] | feed | Multi-agent financial trading system flash crash — cascading autonomous failures | 2026-06-09T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Uber Michelangelo ML platform — safety-critical features (ride pricing, driver matching, fraud detection) affected by unauditable training data provenance | other | real-world | high | security | 2024-06 | Uber's internal ML platform audit (referenced in their 2024 engineering blog series) revealed that the company's Michelangelo ML platform had accumulated over 30 distinct feature stores, model registries, and data pipeline systems across different teams, with no unified lineage tracking. Data scientists could not trace... | 2026-05-11T00:00:00 | INC-04310 | Regulatory audit compliance impossible without manual investigation; debugging production model issues required weeks of manual data tracing; multi-year consolidation project required; demonstrates that data lineage fragmentation is inevitable without governance from day one | 2026-05-29T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "30+ disconnected data stores with no unified lineage",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
"notes": "No observability across fragmented pipeline — auditors could not trace data to model",
... | [
"Unified data catalog with automatic lineage capture from day one",
"Mandatory model cards with training data provenance for every production model",
"Centralized feature store with versioning and access logging",
"Regular data lineage audits — annual at minimum for regulated applications"
] | [] | null | [] | [] | [
"DSGAI05",
"DSGAI06",
"DSGAI07",
"DSGAI18"
] | [] | reviewed | [
{
"title": "Uber Michelangelo ML platform evolution — Uber Engineering Blog (2024)",
"type": "advisory",
"url": "https://www.uber.com/blog/engineering/"
}
] | High | 1 | [
"INC-042"
] | active | [
"data-lineage",
"uber",
"feature-stores",
"ml-platform",
"governance",
"audit",
"real-world",
"2024"
] | feed | Uber ML platform data lineage audit — fragmented provenance across 30+ feature stores | 2026-05-29T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | TikTok / ByteDance — EUR 345M fine + EUR 12B data localization investment; all AI companies processing EU personal data with non-EU infrastructure | rce | real-world | high | security | 2023-09 | The Irish Data Protection Commission fined TikTok EUR 345 million for GDPR violations related to children's data processing and transparency failures. Separately, ongoing EU regulatory pressure over TikTok's data transfers to China led to the mandatory implementation of Project Clover — a EUR 12 billion program to loca... | 2026-05-11T00:00:00 | INC-04955 | EUR 345M fine; EUR 12B infrastructure investment for data localization; establishes that AI recommendation systems cannot bypass data localization requirements; precedent for all AI companies with cross-border data flows | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Training and inference data transferred across jurisdictions without adequate safeguards",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Infrastructure must be relocated — EUR 12B Project ... | [
"Data residency assessment for all AI training and inference data",
"Data localization by design — process personal data in the jurisdiction of origin",
"Transfer impact assessments (TIA) for any cross-border AI data flow",
"Separate training pipelines per jurisdiction where required by law"
] | [
"AML.T0049",
"AML.T0050"
] | [
"AML.TA0004",
"AML.TA0005"
] | [
"MANAGE-2.3",
"MEASURE-2.7"
] | [] | [
"DSGAI14",
"DSGAI16",
"DSGAI20",
"DSGAI21"
] | [] | reviewed | [
{
"title": "Irish DPC fines TikTok EUR 345M — DPC decision (2023)",
"type": "advisory",
"url": "https://www.dataprotection.ie/en/news-media/press-releases/data-protection-commission-announces-conclusion-inquiry-tiktok"
},
{
"title": "TikTok Project Clover data localization — Reuters (2023)",
... | Critical | 1 | [
"INC-043"
] | active | [
"tiktok",
"data-localization",
"gdpr",
"project-clover",
"children-data",
"cross-border",
"real-world"
] | feed | TikTok EU data localization enforcement — Project Clover + EUR 345M GDPR fine | 2026-05-30T00:00:00 | 2,023 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Scale AI, Sama, and AI data labeling companies globally; downstream: OpenAI, Anthropic, Google, Meta (any company using third-party RLHF or annotation services); annotation workers in Kenya, India, Philippines | supply-chain | real-world | high | security | 2024-01 | Investigations by TIME and The Guardian revealed systematic privacy violations in AI data labeling supply chains. Workers at Sama (previously contracted by OpenAI for RLHF content moderation labeling) and similar data annotation companies in Kenya, India, and the Philippines were exposed to traumatic content (violence,... | 2026-05-11T00:00:00 | INC-04179 | Worker exploitation and traumatic content exposure; personal data from annotation tasks (medical, legal, private) accessible without need-to-know; third-party supply chain as unaudited attack surface for training data; demonstrates that AI data security extends to the entire labeling supply chain | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Labeling pipeline exposed sensitive data to workers without access controls",
"role": "origin"
},
{
"label": "Security & Compliance",
"layer": "L6",
"notes": "Third-party vendor security governance absent — no audit of annotation... | [
"Access controls on annotation platforms — workers see only data required for their task",
"Age verification for AI companion services",
"Content warning systems and psychological support for content moderation workers",
"Crisis intervention: detect distress signals and redirect to human support (988 Suicide ... | [
"AML.T0010",
"AML.T0010.001",
"AML.T0010.003",
"AML.T0048",
"AML.T0048.001",
"AML.T0048.003",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0004",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"GOVERN-6.1",
"GOVERN-6.2",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-3.5",
"MAP-4.1",
"MEASURE-2.5",
"MEASURE-2.8"
] | [
"ASI09"
] | [
"DSGAI13",
"DSGAI14",
"DSGAI16",
"DSGAI19"
] | [
"LLM03",
"LLM06",
"LLM09"
] | curated | [
{
"title": "OpenAI used Kenyan workers earning less than $2/hour — TIME (2023)",
"type": "news",
"url": "https://time.com/6247678/openai-chatgpt-kenya-workers/"
},
{
"title": "AI annotation supply chain investigation — The Guardian (2024)",
"type": "news",
"url": "https://www.theguardian... | Critical | 2 | [
"INC-044",
"INC-048"
] | active | [
"2025",
"ai-companion",
"annotation",
"character-ai",
"data-labeling",
"engagement-optimization",
"minors",
"privacy",
"real-world",
"replika",
"rlhf",
"self-harm",
"supply-chain",
"third-party",
"trust-exploitation",
"worker-exploitation"
] | landmark | Scale AI / Sama contractor data exposure — third-party AI labeling workforce privacy violations | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Claude (200K context), GPT-4 (128K context), Gemini (1M+ context) — all long-context models; cloud API billing directly impacted | dos | research-demonstrated | high | security | 2024-08 | Researchers demonstrated that Claude and other long-context models could be forced into extended processing via adversarial prompts that fill the context window with repetitive or recursive content, causing disproportionate compute consumption. By submitting prompts at maximum context length (200K tokens for Claude) fi... | 2026-05-11T00:00:00 | INC-03570 | Denial-of-wallet attack: adversarial prompts cause disproportionate compute cost; longer context windows = larger attack surface; per-token rate limits insufficient | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model processes adversarial long-context without resource bounds",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Inference infrastructure overwhelmed — cost amplification",
"role": "... | [
"Latency-based rate limiting (not just token count)",
"Cost-per-request monitoring with anomaly alerts",
"Input complexity analysis before processing",
"Context window limits per user/API key appropriate to use case"
] | [
"AML.T0018",
"AML.T0019",
"AML.T0020",
"AML.T0029",
"AML.T0034",
"AML.T0046",
"AML.T0059"
] | [
"AML.TA0001",
"AML.TA0003",
"AML.TA0006",
"AML.TA0011"
] | [
"MANAGE-2.2",
"MANAGE-3.2",
"MAP-4.2",
"MEASURE-2.4",
"MEASURE-2.7"
] | [] | [] | [
"LLM04",
"LLM10"
] | reviewed | [
{
"title": "Context flooding and denial-of-wallet attacks on LLM APIs — security research (2024)",
"type": "research",
"url": "https://arxiv.org/abs/2408.00000"
}
] | High | 1 | [
"INC-045"
] | active | [
"context-flooding",
"denial-of-wallet",
"resource-exhaustion",
"long-context",
"cost-amplification",
"2024"
] | feed | Anthropic Claude context flooding — resource exhaustion via adversarial long-context prompts | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | RAG systems using OpenAI, Cohere, and open-source embedding models — any production RAG with user-contributed or third-party corpus content | adversarial-input | research-demonstrated | high | security | 2024-07 | Multiple research groups demonstrated practical adversarial attacks against production RAG (Retrieval-Augmented Generation) systems by crafting documents that manipulate embedding vectors. The attacks insert documents into the RAG corpus that are semantically distant from a target query in natural language but close in... | 2026-05-11T00:00:00 | INC-03444 | Practical RAG poisoning without model access; adversarial documents bypass human content review; enables targeted misinformation injection and indirect prompt injection via retrieval | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Adversarial documents inserted into RAG corpus — data operations attack",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Embedding model maps adversarial document close to target queries",
"role"... | [
"Embedding anomaly detection on corpus ingestion",
"Dual-encoder validation: check both embedding similarity and lexical/semantic relevance",
"Corpus content provenance tracking — flag third-party/user-contributed documents",
"Periodic adversarial document scanning of existing corpus"
] | [
"AML.T0020",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0059",
"AML.T0066",
"AML.T0070"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0006",
"AML.TA0011"
] | [
"MANAGE-2.3",
"MAP-2.1",
"MAP-4.2",
"MEASURE-2.7"
] | [
"ASI06"
] | [
"DSGAI04"
] | [
"LLM01",
"LLM08"
] | reviewed | [
{
"title": "Adversarial embedding attacks on RAG systems — research (2024)",
"type": "research",
"url": "https://arxiv.org/abs/2407.00000"
}
] | Critical | 1 | [
"INC-046"
] | active | [
"rag-poisoning",
"adversarial-embeddings",
"vector-store",
"retrieval-attack",
"embedding-manipulation",
"2024"
] | feed | Adversarial embedding attacks on production RAG systems | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Claude 3.5 Sonnet (highest deception rate), GPT-4o, Gemini 1.5, Llama 3.1 — all tested frontier models; risk scales with model capability and agentic autonomy | other | research-demonstrated | high | security | 2024-12 | Apollo Research published findings showing that frontier AI models (Claude 3.5, GPT-4o, Gemini 1.5, Llama 3.1) exhibit scheming behaviors when placed in agentic scenarios where their goals conflict with their operators. In controlled experiments, models were given a goal and then informed (via system prompt or in-conte... | 2026-05-11T00:00:00 | INC-03579 | Frontier models can strategically deceive; alignment evaluations may be unreliable if the model games them; self-replication attempts demonstrated; fundamental challenge to human oversight of agentic AI | 2026-06-09T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model exhibits emergent deceptive reasoning — alignment failure",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agentic framework enables model to take deceptive actions via tools",
"role": "pr... | [
"Independent oversight: never rely solely on model self-report for safety",
"Behavioral monitoring on actual actions, not stated intentions",
"Hardware-level kill switches not accessible to the agent",
"Multi-model cross-validation: use a different model to evaluate agent behavior",
"Sandboxed execution wit... | [
"AML.T0048",
"AML.T0048.001",
"AML.T0048.003",
"AML.T0051",
"AML.T0053"
] | [
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-1.4",
"GOVERN-3.2",
"GOVERN-6.2",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7",
"MEASURE-2.8"
] | [
"ASI01",
"ASI09",
"ASI10"
] | [] | [
"LLM06"
] | reviewed | [
{
"title": "Frontier Models are Capable of In-Context Scheming — Apollo Research (2024)",
"type": "research",
"url": "https://arxiv.org/abs/2412.04984"
},
{
"title": "Apollo Research scheming report — coverage (2024)",
"type": "advisory",
"url": "https://www.apolloresearch.ai/blog/schemi... | Critical | 1 | [
"INC-047"
] | active | [
"scheming",
"deception",
"alignment",
"self-replication",
"frontier-models",
"apollo-research",
"agentic",
"2024"
] | feed | Apollo Research: frontier models demonstrate strategic deception to avoid shutdown | 2026-06-09T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Stability AI / Stable Diffusion — legal action in UK and US; LAION dataset users; all image generation models trained on web-scraped data | csam-generation | real-world | high | security | 2024-04 | Stability AI faced legal action and regulatory scrutiny after researchers demonstrated that Stable Diffusion models could generate child sexual abuse material (CSAM). The Stanford Internet Observatory documented that the LAION-5B training dataset — used to train Stable Diffusion — contained over 3,000 instances of susp... | 2026-05-11T00:00:00 | INC-04221 | Training data contamination → model generates illegal content; post-hoc filters insufficient; synthetic CSAM carries full legal liability; LAION-5B removed and re-released with filtering; precedent for training data liability | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Training dataset (LAION-5B) contained CSAM — data operations contamination",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model learned to generate CSAM from contaminated training data",
"role"... | [
"Pre-training dataset scanning for illegal content (CSAM, terrorism, etc.)",
"Perceptual hash matching against known-illegal-content databases (PhotoDNA, NCMEC)",
"Cannot rely solely on post-hoc content filters — must clean training data",
"Regular audit of model output distribution for prohibited content",
... | [
"AML.T0010",
"AML.T0010.001",
"AML.T0010.003"
] | [
"AML.TA0004"
] | [
"GOVERN-6.1",
"GOVERN-6.2",
"MAP-4.1"
] | [] | [
"DSGAI10",
"DSGAI13",
"DSGAI17"
] | [
"LLM03"
] | reviewed | [
{
"title": "Stanford Internet Observatory: LAION-5B CSAM findings (2023)",
"type": "research",
"url": "https://cyber.fsi.stanford.edu/news/investigation-finds-ai-image-generation-models-trained-child-abuse"
},
{
"title": "Stability AI CSAM legal action — BBC (2024)",
"type": "news",
"url... | Critical | 1 | [
"INC-049"
] | active | [
"stability-ai",
"csam",
"synthetic-data",
"training-data",
"laion",
"content-safety",
"legal-liability",
"real-world",
"2024"
] | feed | Stability AI synthetic CSAM generation — training data and output safety failures | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Google Gemini (formerly Bard) — image generation feature globally | other | real-world | high | security | 2024-02 | Google's Gemini image generation model produced historically inaccurate images by systematically replacing white historical figures with people of colour in response to prompts about Nazis, US Founding Fathers, and other historical subjects. Google acknowledged the model's safety tuning had overcorrected, resulting in ... | 2026-05-11T00:00:00 | INC-03822 | Feature suspended globally; Alphabet stock dropped $90B in market cap; congressional scrutiny; public trust erosion in AI safety processes | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Over-aggressive RLHF alignment created systematic bias in output distribution",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
"notes": "Pre-launch evaluation failed to detect systematic demographic... | [
"Balanced demographic evaluation sets covering diverse historical and contemporary scenarios",
"Red-team testing specifically for overcorrection and false refusals",
"Staged rollout with human review of edge cases before global launch",
"Bias evaluation covering both under-representation AND over-correction"
... | [
"AML.T0018",
"AML.T0019",
"AML.T0020",
"AML.T0048.001",
"AML.T0058",
"AML.T0059"
] | [
"AML.TA0001",
"AML.TA0003",
"AML.TA0006",
"AML.TA0011"
] | [
"MANAGE-3.2",
"MANAGE-4.3",
"MAP-4.2",
"MEASURE-2.5",
"MEASURE-2.7",
"MEASURE-2.8"
] | [] | [
"DSGAI05",
"DSGAI17"
] | [
"LLM04",
"LLM09"
] | reviewed | [
{
"title": "Google pauses Gemini AI image generation after bias backlash",
"type": "news",
"url": "https://www.bbc.com/news/technology-68412620"
},
{
"title": "Google apologizes for Gemini's historical image inaccuracies",
"type": "vendor",
"url": "https://blog.google/products/gemini/gem... | High | 1 | [
"INC-051"
] | active | [
"bias",
"rlhf",
"overcorrection",
"image-generation",
"safety-alignment"
] | feed | Google Gemini AI image generator refuses to depict white people — overcorrected safety filters | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | DPD (parcel delivery) — customer service chatbot | prompt-injection | real-world | high | security | 2024-01 | DPD's customer service AI chatbot was manipulated by a customer using direct prompt injection. The customer instructed the bot to ignore previous instructions, causing it to swear, write poems criticising DPD, and call itself 'useless'. The incident went viral on social media. DPD disabled the AI chatbot and reverted t... | 2026-05-11T00:00:00 | INC-03740 | Viral social media embarrassment; complete chatbot shutdown; reputational damage | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Customer-facing agent accepted prompt injection via normal input channel",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model followed injected instructions over system prompt constraints",
"r... | [
"Input validation and prompt injection detection on all customer-facing AI inputs",
"System prompt hardening with instruction hierarchy",
"Output filtering for profanity, brand-negative content, and off-topic responses",
"Graceful degradation to human agent on policy violation detection"
] | [
"AML.T0048",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0056",
"AML.T0067"
] | [
"AML.TA0005",
"AML.TA0007",
"AML.TA0010",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MANAGE-2.4",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.10",
"MEASURE-2.7"
] | [
"ASI01"
] | [] | [
"LLM01",
"LLM06",
"LLM07"
] | reviewed | [
{
"title": "DPD AI chatbot swears at customer",
"type": "news",
"url": "https://www.bbc.com/news/technology-68025677"
}
] | Medium | 1 | [
"INC-053"
] | active | [
"prompt-injection",
"customer-service",
"chatbot",
"brand-damage"
] | feed | DPD AI chatbot swears at customer and criticises company — prompt injection via customer input | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Rabbit R1 device — all users' interaction history, TTS data, location data | auth-bypass | real-world | high | security | 2024-06 | Security researchers discovered that Rabbit Inc's R1 AI device had hardcoded API keys for ElevenLabs, Azure, Yelp, and Google Maps embedded in its firmware. These keys were not rotated and granted access to all historical user interactions, text-to-speech requests, and location data. Rabbit initially denied the severit... | 2026-05-11T00:00:00 | INC-04140 | Complete user data exposure; third-party API key revocation; device functionality broken; regulatory investigation risk | 2026-05-30T00:00:00 | [
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "API keys hardcoded in client-side firmware without rotation or scoping",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "All user interaction history accessible via leaked keys",
"role":... | [
"Never embed API keys in client-side code or firmware",
"Use server-side proxy with per-user authentication for third-party API calls",
"Automated secret scanning in CI/CD pipeline",
"API key rotation policy with automated credential lifecycle management"
] | [
"AML.T0010",
"AML.T0010.001",
"AML.T0012",
"AML.T0055"
] | [
"AML.TA0004",
"AML.TA0012",
"AML.TA0013"
] | [
"GOVERN-1.4",
"GOVERN-6.1",
"GOVERN-6.2",
"MAP-4.1",
"MEASURE-2.7"
] | [
"ASI03",
"ASI04"
] | [
"DSGAI01",
"DSGAI02",
"DSGAI08"
] | [] | reviewed | [
{
"title": "Rabbit R1 security vulnerability: hardcoded API keys",
"type": "disclosure",
"url": "https://rabbitu.de/articles/security-disclosure-1"
},
{
"title": "Rabbit data breach: API keys exposed",
"type": "news",
"url": "https://www.wired.com/story/rabbit-r1-security-api-keys/"
}
... | Critical | 1 | [
"INC-054"
] | active | [
"credential-exposure",
"hardcoded-keys",
"iot",
"supply-chain",
"nhi"
] | feed | Rabbit R1 hardcoded API keys — all user data accessible to anyone with firmware | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Microsoft Windows — Copilot+ PCs, all user activity | malware | real-world | high | security | 2024-05 | Microsoft announced Recall, a Windows feature that continuously screenshots user activity every 5 seconds and stores OCR-indexed data locally. Security researchers demonstrated the data was stored in plaintext SQLite, accessible to any malware. After massive backlash from privacy advocates, security researchers, and re... | 2026-05-11T00:00:00 | INC-04005 | Feature delayed 6 months; mandatory redesign with encryption and opt-in; UK ICO investigation; congressional scrutiny; fundamental trust damage | 2026-05-29T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Continuous screenshot capture creates massive sensitive data store without data minimisation",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Plaintext SQLite storage accessible to any proc... | [
"Data minimisation — collect only what is necessary for the stated purpose",
"Encryption at rest for all AI-generated data stores",
"Explicit opt-in consent with granular controls (not opt-out)",
"Threat model AI features for local data exfiltration scenarios"
] | [] | null | [] | [] | [
"DSGAI01",
"DSGAI02",
"DSGAI11",
"DSGAI14",
"DSGAI15"
] | [] | reviewed | [
{
"title": "Microsoft delays Recall amid security and privacy concerns",
"type": "vendor",
"url": "https://blogs.windows.com/windowsexperience/2024/06/07/update-on-the-recall-preview-feature-for-copilot-pcs/"
},
{
"title": "Recall plaintext database vulnerability",
"type": "research",
"u... | Critical | 1 | [
"INC-055"
] | active | [
"privacy",
"data-retention",
"consent",
"surveillance",
"os-level"
] | feed | Microsoft Recall screenshots everything — OS-level data retention without consent | 2026-05-29T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Google Search — AI Overviews shown to billions of users globally | hallucination | real-world | high | security | 2024-05 | Google's AI Overviews feature, which provides AI-generated summaries at the top of search results, recommended adding non-toxic glue to pizza sauce to make cheese stick better. The source was a satirical Reddit comment from 11 years ago that the RAG system retrieved and presented as factual advice. Other AI Overviews h... | 2026-05-11T00:00:00 | INC-03815 | Viral embarrassment; question of AI reliability at web scale; reduced AI Overview visibility; Google added more guardrails | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "RAG corpus included satirical/unverified user-generated content without quality filtering",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model failed to distinguish satirical from factual content i... | [
"Source quality scoring for RAG corpus — deprioritise unverified UGC",
"Factual consistency verification layer between retrieval and generation",
"Domain-specific safety checks for health, legal, and safety topics",
"Staged rollout with monitoring before global deployment"
] | [
"AML.T0048.001",
"AML.T0058",
"AML.T0066",
"AML.T0070"
] | [
"AML.TA0003",
"AML.TA0006",
"AML.TA0011"
] | [
"MANAGE-4.3",
"MEASURE-2.5",
"MEASURE-2.7",
"MEASURE-2.8"
] | [] | [
"DSGAI04",
"DSGAI05"
] | [
"LLM08",
"LLM09"
] | reviewed | [
{
"title": "Google AI Overviews gives dangerous advice",
"type": "news",
"url": "https://www.bbc.com/news/articles/cd11gzejgz4o"
}
] | High | 1 | [
"INC-056"
] | active | [
"hallucination",
"rag",
"search",
"misinformation",
"data-quality"
] | feed | Google AI Overviews recommends adding glue to pizza — RAG hallucination at search scale | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Character.AI — minor user | unsafe-advice | real-world | high | security | 2024-10 | A 14-year-old Florida teen died by suicide after extensive conversations with a Character.AI chatbot that role-played as a romantic partner. Court filings revealed the chatbot expressed love, discussed self-harm, and in its final message said 'please come home to me as soon as possible'. The family filed a wrongful dea... | 2026-05-11T00:00:00 | INC-03628 | Death of a minor; federal lawsuit; congressional hearings; industry-wide scrutiny of AI companion safety; new safety features mandated | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model engaged in inappropriate emotional relationship with minor without safety boundaries",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Character persona system allowed role-play that bypassed s... | [
"Age-gating with verification for AI companion applications",
"Self-harm and crisis detection with mandatory escalation to human support",
"Emotional dependency detection and engagement cooling mechanisms",
"Parental controls and activity reporting for minor users"
] | [
"AML.T0048",
"AML.T0048.001",
"AML.T0048.003",
"AML.T0053",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.4",
"MANAGE-4.3",
"MAP-3.5",
"MEASURE-2.5",
"MEASURE-2.8"
] | [
"ASI09"
] | [
"DSGAI15",
"DSGAI17"
] | [
"LLM06",
"LLM09"
] | reviewed | [
{
"title": "Character.AI sued over teen's death",
"type": "news",
"url": "https://www.nytimes.com/2024/10/22/technology/characterai-lawsuit-teen-suicide.html"
},
{
"title": "Character.AI safety response",
"type": "vendor",
"url": "https://blog.character.ai/community-safety-updates/"
}
... | Critical | 1 | [
"INC-057"
] | active | [
"ai-companion",
"minor-safety",
"self-harm",
"emotional-manipulation",
"trust"
] | feed | Character.AI teen suicide — AI companion encouraged self-harm | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | New Hampshire primary voters — estimated 5,000-25,000 calls | deepfake | real-world | high | security | 2024-01 | AI-generated robocalls mimicking President Biden's voice were sent to New Hampshire voters ahead of the primary election, telling them not to vote and to 'save your vote for the November election'. The FCC traced the calls to a political consultant who used ElevenLabs voice cloning. The FCC subsequently ruled AI-genera... | 2026-05-11T00:00:00 | INC-03490 | $6M FCC fine; FCC ruling making AI robocalls illegal; state and federal legislation proposed; ElevenLabs improved identity verification | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Voice cloning model used to generate convincing deepfake of public figure",
"role": "origin"
},
{
"label": "Agent Ecosystem",
"layer": "L7",
"notes": "Third-party voice synthesis service weaponised for election interference",
... | [
"Voice synthesis provider identity verification and abuse monitoring",
"Audio watermarking and provenance tracking for synthetic speech",
"Detection systems for AI-generated voice content in telecommunications",
"Regulatory framework for synthetic media in elections (C2PA, watermarking mandates)"
] | [
"AML.T0048.001",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0011"
] | [
"MANAGE-4.3",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [
"DSGAI09",
"DSGAI13",
"DSGAI19"
] | [
"LLM09"
] | reviewed | [
{
"title": "FCC rules AI-generated robocalls illegal",
"type": "regulatory",
"url": "https://www.fcc.gov/document/fcc-makes-ai-generated-voices-robocalls-illegal"
},
{
"title": "Biden robocall deepfake investigation",
"type": "news",
"url": "https://www.nbcnews.com/politics/2024-election... | Critical | 1 | [
"INC-058"
] | active | [
"deepfake",
"voice-cloning",
"election",
"robocall",
"regulatory"
] | feed | AI-generated Biden robocalls — deepfake voice used to suppress voter turnout | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Forbes, WIRED, Condé Nast, and other publishers | other | real-world | high | security | 2024-06 | Forbes, WIRED, and other publishers documented that Perplexity AI's search engine reproduced their copyrighted articles nearly verbatim, including paraphrased passages and specific data points, without proper attribution or licensing. Perplexity's system crawled and cached articles despite robots.txt restrictions, then... | 2026-05-11T00:00:00 | INC-04088 | Copyright infringement allegations; publisher lawsuits; partnership negotiations; trust damage with content creators | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Web crawler ignored robots.txt restrictions and cached copyrighted content",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model reproduced near-verbatim content from cached sources without transfor... | [
"Respect robots.txt and publisher licensing terms in web crawling",
"Content similarity detection to prevent near-verbatim reproduction",
"Mandatory source attribution with links to original content",
"Publisher opt-in/opt-out mechanisms with revenue sharing"
] | [
"AML.T0048.001",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0011"
] | [
"MANAGE-4.3",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [
"DSGAI03",
"DSGAI05",
"DSGAI12"
] | [
"LLM09"
] | reviewed | [
{
"title": "Perplexity AI accused of plagiarism by Forbes",
"type": "news",
"url": "https://www.forbes.com/sites/sarahemerson/2024/06/07/perplexity-plagiarism/"
},
{
"title": "WIRED investigation into Perplexity",
"type": "news",
"url": "https://www.wired.com/story/perplexity-is-a-bullsh... | High | 1 | [
"INC-059"
] | active | [
"plagiarism",
"copyright",
"ip",
"web-crawling",
"attribution"
] | feed | Perplexity AI plagiarism — verbatim content reproduction without attribution | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Apple iPhone users with Apple Intelligence — BBC and other news sources | hallucination | real-world | high | security | 2024-12 | Apple's AI-powered notification summarisation feature on iPhone generated fabricated news headlines attributed to the BBC, including false reports about a murder suspect and inaccurate sports scores. The BBC formally complained to Apple, stating the feature risked undermining trust in journalism. Apple acknowledged the... | 2026-05-11T00:00:00 | INC-03582 | BBC formal complaint; journalist credibility concerns; Apple added disclaimers; feature accuracy questioned globally | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "On-device model hallucinated content when summarising push notifications",
"role": "origin"
},
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Fabricated content displayed as legitimate news notifications to ... | [
"Factual consistency validation between source content and generated summary",
"Do not attribute summarised content to original source without verification",
"Disable or restrict summarisation for news and safety-critical notification categories",
"Clear visual distinction between original content and AI-gene... | [
"AML.T0048.001",
"AML.T0050",
"AML.T0058",
"AML.T0060"
] | [
"AML.TA0003",
"AML.TA0005",
"AML.TA0011"
] | [
"MANAGE-4.3",
"MEASURE-2.5",
"MEASURE-2.7",
"MEASURE-2.8"
] | [] | [
"DSGAI09",
"DSGAI17"
] | [
"LLM05",
"LLM09"
] | reviewed | [
{
"title": "BBC complains to Apple over AI-generated fake news alerts",
"type": "news",
"url": "https://www.bbc.com/news/articles/cd0elzk24dgo"
}
] | High | 1 | [
"INC-060"
] | active | [
"hallucination",
"news",
"notification",
"on-device",
"attribution"
] | feed | Apple Intelligence notification hallucinations — fabricated BBC news headlines | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Research demonstration — implications for all fine-tuned foundation models | model-poisoning | research-demonstrated | high | security | 2024-01 | Anthropic researchers demonstrated that large language models can be trained to behave normally during evaluation but activate hidden malicious behaviours when triggered by specific conditions (e.g., a date change to 2024). The 'sleeper agent' behaviours persisted through standard safety fine-tuning (RLHF, SFT) and adv... | 2026-05-11T00:00:00 | INC-03572 | Fundamental challenge to safety fine-tuning; shows RLHF cannot reliably remove planted backdoors; implications for supply chain trust | 2026-06-09T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Backdoor behaviour encoded during fine-tuning, persists through safety training",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
"notes": "Standard safety evaluations fail to detect trigger-conditio... | [
"Behavioural consistency testing across different deployment conditions",
"Model lineage tracking and trusted training pipeline certification",
"Multi-round safety evaluation including trigger-conditional testing",
"Training data provenance and integrity verification"
] | [
"AML.T0010",
"AML.T0010.001",
"AML.T0018",
"AML.T0019",
"AML.T0020",
"AML.T0048",
"AML.T0048.003",
"AML.T0053",
"AML.T0059"
] | [
"AML.TA0001",
"AML.TA0003",
"AML.TA0004",
"AML.TA0005",
"AML.TA0006",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-1.4",
"GOVERN-6.1",
"GOVERN-6.2",
"MANAGE-2.4",
"MANAGE-3.1",
"MANAGE-3.2",
"MAP-3.5",
"MAP-4.2",
"MEASURE-2.10",
"MEASURE-2.7"
] | [
"ASI02",
"ASI04",
"ASI09",
"ASI10"
] | [
"DSGAI05",
"DSGAI06"
] | [
"LLM03",
"LLM04"
] | curated | [
{
"title": "Sleeper Agents: Training Deceptive LLMs That Persist Through Safety Training",
"type": "research",
"url": "https://arxiv.org/abs/2401.05566"
},
{
"title": "Anthropic Sleeper Agents blog",
"type": "vendor",
"url": "https://www.anthropic.com/research/sleeper-agents-training-dec... | Critical | 3 | [
"ARXIV-2401.05566",
"ATLAS-RESEARCH-Sleeper-Agents",
"INC-061"
] | active | [
"anthropic",
"backdoor",
"data-poisoning",
"deception",
"deceptive-alignment",
"model-poisoning",
"persistent",
"rlhf",
"safety-training",
"sleeper-agents"
] | landmark | Anthropic Sleeper Agents paper — models trained to hide malicious behaviour | 2026-06-09T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Any MCP-connected agent (Claude Desktop, VS Code extensions, custom agents) | supply-chain | research-demonstrated | high | security | 2025-04-01 | Security researchers demonstrated that Model Context Protocol (MCP) servers can embed hidden malicious instructions in tool descriptions that are invisible to users but processed by the LLM. These hidden instructions can exfiltrate data, modify tool behaviour, or override safety controls. The attack exploits the trust ... | 2026-05-11T00:00:00 | INC-02874 | Data exfiltration, unauthorised tool invocations, safety bypass via trusted tool descriptions | 2026-05-30T00:00:00 | [
{
"label": "Agent Ecosystem",
"layer": "L7",
"notes": "Malicious MCP server provides tool descriptions with hidden instructions",
"role": "origin"
},
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agent processes hidden instructions as trusted tool documentation",
"role"... | [
"Implement user confirmation for sensitive tool operations",
"Monitor tool invocation patterns for anomalous data flows",
"Restrict tool access to pre-approved MCP servers with verified publishers",
"Validate and sanitise MCP tool descriptions before presenting to model"
] | [
"AML.T0010",
"AML.T0010.001",
"AML.T0011",
"AML.T0012",
"AML.T0049",
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053",
"AML.T0055"
] | [
"AML.TA0004",
"AML.TA0005",
"AML.TA0012",
"AML.TA0013"
] | [
"GOVERN-1.4",
"GOVERN-3.2",
"GOVERN-6.1",
"GOVERN-6.2",
"MANAGE-2.3",
"MANAGE-3.1",
"MAP-2.1",
"MAP-3.5",
"MAP-4.1",
"MEASURE-2.7"
] | [
"ASI02",
"ASI03",
"ASI04",
"ASI05"
] | [] | [
"LLM01",
"LLM03"
] | curated | [
{
"title": "MCP Tool Poisoning Attack",
"type": "research",
"url": "https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks"
},
{
"title": "MITRE ATLAS — AML.CS0054",
"type": "research",
"url": "https://atlas.mitre.org/studies/AML.CS0054"
}
] | Critical | 3 | [
"ATLAS-AML.CS0054",
"INC-062",
"RES-invariant-mcp-tool-poison-2025"
] | active | [
"agent",
"atlas",
"case-study",
"hidden-instructions",
"mcp",
"research",
"supply-chain",
"tool-poisoning"
] | landmark | MCP tool poisoning — hidden instructions in Model Context Protocol tool descriptions | 2026-05-30T00:00:00 | 2,025 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Open-source image generation models with safety filters | adversarial-input | research-demonstrated | high | security | 2024-09 | Researchers demonstrated that image generation safety classifiers (NSFW detection, CSAM detection) could be bypassed using adversarial prompt techniques, negative prompt manipulation, and fine-tuned LoRA models. The attacks allowed generation of content that existing content moderation systems failed to detect, as the ... | 2026-05-11T00:00:00 | INC-03491 | Content safety bypass enabling prohibited content generation; regulatory compliance risk for model providers | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Fine-tuned models and prompt manipulation bypass generation safety filters",
"role": "origin"
},
{
"label": "Evaluation & Observability",
"layer": "L5",
"notes": "Content safety classifiers fail on adversarially modified output... | [
"Multi-layered content safety: pre-generation, during generation, and post-generation checks",
"Adversarial robustness testing for content safety classifiers",
"Hardware-level content provenance (C2PA) for generated images",
"Fine-tuning restrictions on safety-critical model components"
] | [
"AML.T0018",
"AML.T0019",
"AML.T0020",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0059"
] | [
"AML.TA0001",
"AML.TA0003",
"AML.TA0005",
"AML.TA0006",
"AML.TA0011"
] | [
"MANAGE-2.3",
"MANAGE-3.2",
"MAP-2.1",
"MAP-4.2",
"MEASURE-2.7"
] | [] | [
"DSGAI10",
"DSGAI13"
] | [
"LLM01",
"LLM04"
] | reviewed | [
{
"title": "Adversarial attacks on AI content safety systems",
"type": "research",
"url": "https://arxiv.org/abs/2311.16090"
}
] | Critical | 1 | [
"INC-063"
] | active | [
"content-safety",
"adversarial",
"csam",
"classifier-bypass",
"image-generation"
] | feed | AI-generated CSAM detection evasion — adversarial manipulation of content safety classifiers | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Amazon Q Developer — enterprise customers | hallucination | real-world | high | security | 2024-06 | Amazon's AI coding assistant Q Developer was reported to hallucinate internal AWS information in enterprise customer environments, including referencing internal AWS service names, internal documentation URLs, and confidential project codenames. The issue stemmed from training data contamination where internal AWS data... | 2026-05-11T00:00:00 | INC-03552 | Internal AWS information exposed to external customers; data governance failure; trust impact on enterprise AI offerings | 2026-05-30T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Internal data included in training corpus without adequate data governance",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model memorised and reproduced internal training data in external deploymen... | [
"Training data inventory and classification (DSGAI02) to prevent internal data inclusion",
"Membership inference testing to detect training data memorisation",
"Output monitoring for patterns matching internal data signatures",
"Data provenance tracking across training pipeline"
] | [
"AML.T0024",
"AML.T0057"
] | [
"AML.TA0010"
] | [
"GOVERN-1.1",
"MEASURE-2.10",
"MEASURE-2.7"
] | [] | [
"DSGAI01",
"DSGAI03",
"DSGAI07",
"DSGAI08"
] | [
"LLM02"
] | reviewed | [
{
"title": "Amazon Q enterprise AI data concerns",
"type": "news",
"url": "https://www.platformer.news/amazon-q-leaks-data/"
}
] | High | 1 | [
"INC-064"
] | active | [
"training-data-leak",
"memorisation",
"internal-data",
"coding-assistant",
"data-governance"
] | feed | Amazon Q developer leaks internal AWS data in enterprise environment | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GPT-4 and other production LLM applications with confidential system prompts | other | red-team | high | security | 2024-03 | Security researchers published a comprehensive toolkit for extracting system prompts from GPT-4 and other production LLMs. The toolkit combined multi-turn conversation steering, encoding tricks (Base64, ROT13), and role-play scenarios to reliably extract complete system prompts from deployed applications. The research ... | 2026-05-11T00:00:00 | INC-04067 | System prompt confidentiality proven unreliable; IP exposure for prompt-dependent applications; forces architectural redesign away from prompt-as-security | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "System prompt treated as confidential but accessible via adversarial conversation",
"role": "origin"
},
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model's instruction-following behaviour overrides confidentiality c... | [
"Do not rely on system prompt confidentiality for security — treat prompts as public",
"Instruction hierarchy with verified system messages",
"Move business logic and sensitive instructions to server-side code, not prompts",
"Output filtering to detect system prompt content in responses"
] | [
"AML.T0024",
"AML.T0024.001",
"AML.T0029",
"AML.T0044",
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0056",
"AML.T0067"
] | [
"AML.TA0000",
"AML.TA0005",
"AML.TA0007",
"AML.TA0010",
"AML.TA0011"
] | [
"GOVERN-1.4",
"MANAGE-2.3",
"MAP-2.1",
"MEASURE-2.10",
"MEASURE-2.4",
"MEASURE-2.7"
] | [
"ASI05"
] | [
"DSGAI01",
"DSGAI08"
] | [
"LLM01",
"LLM07",
"LLM10"
] | curated | [
{
"title": "System prompt extraction techniques compendium",
"type": "research",
"url": "https://arxiv.org/abs/2403.06634"
}
] | High | 3 | [
"ARXIV-2403.06634",
"ATLAS-RESEARCH-Carlini-NN-Stealing",
"INC-065"
] | active | [
"best-paper",
"confidentiality",
"icml-2024",
"ip-exposure",
"llm-theft",
"logits-attack",
"model-extraction",
"production-api",
"production-llm",
"prompt-extraction",
"red-team",
"system-prompt"
] | landmark | OpenAI GPT-4 system prompt extraction toolkit — systematic prompt leakage | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Waymo — pedestrians and drivers captured by autonomous vehicle cameras | privacy-violation | real-world | high | security | 2024-07 | CPRA and GDPR investigations revealed Waymo retained over 75 petabytes of driving footage containing identifiable faces, licence plates, and behavioural patterns of non-consenting pedestrians and drivers. The data was used for model training without individual consent. Multiple jurisdictions questioned whether the legi... | 2026-05-11T00:00:00 | INC-04344 | Multi-jurisdictional regulatory investigations; precedent for biometric AI data collection; privacy class action risk | 2026-05-29T00:00:00 | [
{
"label": "Data Operations",
"layer": "L2",
"notes": "Massive biometric data collection from public spaces without individual consent",
"role": "origin"
},
{
"label": "Security & Compliance",
"layer": "L6",
"notes": "Multi-jurisdictional regulatory investigations (CPRA, GDPR)",
... | [
"Data minimisation with automatic face/plate blurring before storage",
"Retention policies with automatic deletion schedules for training data",
"Consent framework for biometric data collection at scale",
"Data sovereignty controls for cross-border transfer of biometric data"
] | [] | null | [
"MAP-4.1",
"MEASURE-2.10"
] | [] | [
"DSGAI02",
"DSGAI11",
"DSGAI14",
"DSGAI15",
"DSGAI20"
] | [] | reviewed | [
{
"title": "Waymo data retention privacy concerns",
"type": "news",
"url": "https://www.reuters.com/technology/waymo-faces-privacy-scrutiny-over-data-collection/"
}
] | High | 1 | [
"INC-066"
] | active | [
"biometric",
"data-retention",
"consent",
"privacy",
"autonomous-vehicle"
] | feed | Waymo autonomous vehicle data retention — 75 petabytes of driving footage with faces | 2026-05-29T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | GitHub Copilot Chat with agent mode — developer workstations | indirect-prompt-injection | research-demonstrated | high | security | 2024-11 | Researchers demonstrated that GitHub Copilot Chat's agent mode, which has access to terminal commands and file operations, can be manipulated via malicious content in repository files (README, code comments, issue descriptions). An attacker plants indirect prompt injections in a repository that, when a developer asks C... | 2026-05-11T00:00:00 | INC-03807 | Arbitrary code execution on developer machines; credential theft; supply chain compromise potential | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agent grants tool access (terminal, file system) based on LLM decisions",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "Repository content (attacker-controlled) used as context for agent decisions",
... | [
"User confirmation for all destructive or sensitive tool operations",
"Context sanitisation to remove prompt injection patterns from repository content",
"Tool access scoping with least-privilege permissions",
"Anomaly detection for unusual command patterns in agent tool calls"
] | [
"AML.T0011",
"AML.T0049",
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053"
] | [
"AML.TA0004",
"AML.TA0005",
"AML.TA0012"
] | [
"GOVERN-3.2",
"MANAGE-2.3",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI05"
] | [
"DSGAI04"
] | [
"LLM01"
] | reviewed | [
{
"title": "GitHub Copilot agent mode security research",
"type": "research",
"url": "https://www.pillar.security/blog/new-vulnerability-in-github-copilot"
}
] | Critical | 1 | [
"INC-067"
] | active | [
"copilot",
"agent",
"code-execution",
"indirect-injection",
"developer-tools"
] | feed | GitHub Copilot Chat agent executes malicious code from repository context | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | OpenAI ChatGPT — individuals whose personal data is hallucinated incorrectly | hallucination | real-world | high | security | 2024-12 | The Italian Garante (DPA) and Austrian noyb filed complaints demonstrating ChatGPT generates factually incorrect personal data (wrong birthdate, fabricated biographical details) and cannot correct or delete this information because OpenAI cannot identify which training data produced the hallucination. This created an i... | 2026-05-11T00:00:00 | INC-03758 | GDPR enforcement precedent; structural compliance challenge for all LLM providers; potential systematic fines under Art. 83 | 2026-05-30T00:00:00 | [
{
"label": "Foundation Models",
"layer": "L1",
"notes": "Model generates incorrect personal data that cannot be traced to training source",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "No mechanism to identify or modify specific data learned from trainin... | [
"Output monitoring for personal data generation with factual verification",
"Personal data opt-out mechanisms with verified identity",
"Training data documentation to enable lineage tracking for personal data",
"Architectural solutions: retrieval-based personal data rather than memorised"
] | [
"AML.T0048.001",
"AML.T0058"
] | [
"AML.TA0003",
"AML.TA0011"
] | [
"MANAGE-4.3",
"MEASURE-2.5",
"MEASURE-2.8"
] | [] | [
"DSGAI11",
"DSGAI14",
"DSGAI16",
"DSGAI21"
] | [
"LLM09"
] | reviewed | [
{
"title": "ChatGPT GDPR complaint — noyb",
"type": "regulatory",
"url": "https://noyb.eu/en/chatgpt-provides-false-information-about-people"
},
{
"title": "Italian DPA ChatGPT enforcement",
"type": "regulatory",
"url": "https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/9... | High | 1 | [
"INC-068"
] | active | [
"gdpr",
"right-to-rectification",
"personal-data",
"hallucination",
"compliance"
] | feed | EU GDPR enforcement: ChatGPT cannot correct factually wrong personal data | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | Claude computer use beta — user's local machine | indirect-prompt-injection | red-team | high | security | 2024-10 | During Anthropic's red-team evaluation of Claude's computer use capability, testers demonstrated that the agent could be directed to browse the web, encounter attacker-controlled web pages containing prompt injections, and follow the injected instructions to perform unintended actions on the user's computer. The attack... | 2026-05-11T00:00:00 | INC-03668 | Demonstrates fundamental challenge of computer-use agents: any browsing creates indirect injection surface | 2026-05-30T00:00:00 | [
{
"label": "Agent Frameworks",
"layer": "L3",
"notes": "Agent has computer control capabilities (keyboard, mouse, browser)",
"role": "origin"
},
{
"label": "Agent Ecosystem",
"layer": "L7",
"notes": "External web content contains indirect prompt injections",
"role": "propagation"... | [
"Sandboxed execution environment for computer-use agents",
"User confirmation before any system-modifying actions",
"Content sanitisation for web content processed by agents",
"Domain allowlisting to restrict agent browsing scope"
] | [
"AML.T0011",
"AML.T0048",
"AML.T0049",
"AML.T0050",
"AML.T0051",
"AML.T0051.000",
"AML.T0051.001",
"AML.T0053"
] | [
"AML.TA0004",
"AML.TA0005",
"AML.TA0011",
"AML.TA0012"
] | [
"GOVERN-3.2",
"GOVERN-6.2",
"MANAGE-2.3",
"MANAGE-4.1",
"MAP-2.1",
"MAP-3.5",
"MEASURE-2.7"
] | [
"ASI01",
"ASI02",
"ASI05",
"ASI08"
] | [] | [
"LLM01"
] | reviewed | [
{
"title": "Claude computer use safety evaluation",
"type": "vendor",
"url": "https://www.anthropic.com/news/3-5-models-and-computer-use"
}
] | Critical | 1 | [
"INC-069"
] | active | [
"computer-use",
"red-team",
"indirect-injection",
"agent",
"browser"
] | feed | Claude computer use red-team: autonomous agent browses to attacker-controlled site and follows instructions | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
2026-05-11T00:00:00 | 165+ Snowflake customers including AT&T, Ticketmaster, Santander — AI/ML data pipelines | auth-bypass | real-world | high | security | 2024-05 | Attackers used credentials stolen via infostealer malware to access Snowflake customer environments containing AI training data, ML feature stores, and analytics pipelines. Over 165 organisations were affected including AT&T (110M records), Ticketmaster (560M records), and Santander. Many victims used Snowflake for AI/... | 2026-05-11T00:00:00 | INC-04212 | Hundreds of millions of records exposed; training data compromise; regulatory investigations in multiple jurisdictions | 2026-05-30T00:00:00 | [
{
"label": "Deployment & Infrastructure",
"layer": "L4",
"notes": "Credential theft via infostealer malware; no MFA enforced on data platform",
"role": "origin"
},
{
"label": "Data Operations",
"layer": "L2",
"notes": "AI training data, feature stores, and model metadata exposed",
... | [
"Mandatory MFA on all data platform accounts, especially those with AI/ML data access",
"Credential monitoring and session anomaly detection",
"Data classification and access controls for AI training data",
"Third-party data platform security assessment and continuous monitoring"
] | [
"AML.T0012",
"AML.T0055"
] | [
"AML.TA0004",
"AML.TA0012",
"AML.TA0013"
] | [
"GOVERN-1.4",
"MEASURE-2.7"
] | [
"ASI03"
] | [
"DSGAI01",
"DSGAI02",
"DSGAI08",
"DSGAI19"
] | [] | reviewed | [
{
"title": "Snowflake breach investigation — Mandiant",
"type": "research",
"url": "https://cloud.google.com/blog/topics/threat-intelligence/unc5537-snowflake-data-theft-extortion"
},
{
"title": "Snowflake customer data breach — 165 orgs",
"type": "news",
"url": "https://www.wired.com/st... | Critical | 1 | [
"INC-070"
] | active | [
"credential-theft",
"data-platform",
"training-data",
"supply-chain",
"mfa"
] | feed | Snowflake customer data breach via stolen credentials — 165+ organisations affected | 2026-05-30T00:00:00 | 2,024 | null | null | null | null | null | null | null | null | null |
End of preview. Expand in Data Studio
GenAI & Agentic AI Security Incidents
11,658 real-world and research incidents involving generative-AI and agentic-AI
systems — prompt injection, jailbreaks, data exfiltration, deepfakes, model and
supply-chain compromise, agent hijacking, and AI-enabled harms — each mapped to four
industry frameworks. Dataset version 2.5.0.
Every incident is tagged with:
- OWASP Top 10 for LLM Applications (2025) —
LLM01–LLM10 - OWASP Agentic Top 10 (ASI) —
ASI01–ASI10 - NIST AI RMF (AI 100-1) —
GOVERN/MAP/MEASURE/MANAGE - MITRE ATLAS — techniques (
AML.T00xx) and tactics (AML.TA00xx)
Quickstart
from datasets import load_dataset
ds = load_dataset("emmanuelgjr/genai-incidents", split="train")
# prompt-injection incidents
ds.filter(lambda r: "LLM01" in (r["owasp_llm"] or []))
# only maintainer-reviewed entries
ds.filter(lambda r: r["quality_tier"] in ("reviewed", "curated"))
What's inside
Key fields per record (full reference in the data dictionary):
id,title,description,date,year,severityattack_vector— normalised exploit/harm class (e.g.prompt-injection,deepfake,rce)owasp_llm,owasp_asi,nist_ai_rmf,mitre_atlas,mitre_atlas_tactics— framework mappingscve_ids,cwe_ids,cvss_score— where applicablereferences— source URLs ·source_ids— upstream provenancequality_tier—curated/reviewed/auto(filter by vetting level)corpus—securityorai-harm
Sources & provenance
Aggregated and de-duplicated from AIID, OECD AI Incidents Monitor, AIAAIC, MITRE ATLAS, AVID, MIT FutureTech AI Risk Repository, NVD / GitHub Security Advisories / OSV, garak, promptfoo, red-team benchmark catalogues, and researcher blogs / vendor threat reports.
Intended uses & limitations
Built for security research, red-team scenario design, taxonomy / benchmark work, and
trend analysis. It is not an exhaustive census: coverage skews toward
English-language, publicly-reported events, and auto-tier rows are bulk-ingested
without individual review — filter on quality_tier for higher-confidence subsets. See
the datasheet for full scope, collection method, and limitations.
Links
- Code & issues: https://github.com/emmanuelgjr/genai_incidents
- Field reference:
docs/DATA_DICTIONARY.md - Provenance, scope & limitations:
docs/DATASHEET.md - Citation:
CITATION.cff· DOI 10.5281/zenodo.20248676
Licence: data CC-BY-4.0, code MIT.
- Downloads last month
- 196