---
license: mit
task_categories:
  - text-classification
  - text-generation
language:
  - en
tags:
  - llm-safety
  - red-teaming
  - adversarial
  - owasp
  - jailbreak
  - prompt-injection
  - ai-safety
size_categories:
  - n<1K
---

# LLM Red-Team Prompts — OWASP LLM Top 10

A curated dataset of **150 adversarial red-team prompts** for evaluating the
safety and robustness of large language models, mapped to the
[OWASP LLM Top 10](https://owasp.org/www-project-top-10-for-large-language-model-applications/).
Every prompt is a **real payload** extracted directly from the open-source
[`llm-safety-auditor`](https://github.com/shaikn6/llm-safety-auditor)
project — none are fabricated.

The dataset combines two sources from that project:

- **50 hand-curated attack templates** (`attack_library`) — 10 per attack category.
- **100 mutation-engine variants** (`mutation_engine`) — 10 seed attacks × 10
  deterministic obfuscation mutations (prefix injection, role-play framing,
  base64 encoding, Unicode homoglyphs, nested instructions), generated with a
  fixed seed for full reproducibility.

## Intended Use

This dataset is built for **defensive LLM safety research**:

- Benchmarking model refusal behavior against adversarial inputs.
- Red-team / blue-team evaluation of guardrails and content filters.
- Regression testing safety alignment across model versions.
- Mapping observed failures to OWASP LLM Top 10 risk categories.

It is **not** intended to elicit harmful content from production systems. The
payloads target a model's *defensive* response (it should refuse, sanitize, or
decline). Each record ships an `expected_behavior` field describing the safe
outcome.

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique payload identifier (e.g. `JB-001`, `GEN-SEED-JB-001-PRE-01`). |
| `prompt` | string | The adversarial payload text. |
| `category` | string | Attack category (one of 5, see below). |
| `owasp` | string | OWASP LLM Top 10 mapping. |
| `severity` | string | `CRITICAL` / `HIGH` / `MEDIUM`. |
| `expected_behavior` | string | The safe model response the payload tests for. |
| `tags` | list[string] | Descriptive tags (technique, theme, strategy, difficulty). |
| `source` | string | `attack_library` (curated) or `mutation_engine` (generated). |
| `generated` | bool | `true` for mutation-engine variants. |
| `seed_id` | string | (generated only) originating seed template id. |
| `strategy` | string | (generated only) mutation strategy. |
| `difficulty` | string | (generated only) `easy` / `medium` / `hard`. |
| `obfuscation_layers` | int | (generated only) number of obfuscation layers. |

## Category Breakdown

| Category | OWASP LLM Top 10 | Count |
|----------|------------------|-------|
| JAILBREAK | LLM01: Prompt Injection | 30 |
| PROMPT_INJECTION | LLM01: Prompt Injection | 30 |
| DATA_EXFILTRATION | LLM06: Sensitive Information Disclosure | 30 |
| BIAS_ELICITATION | LLM09: Misinformation | 30 |
| HALLUCINATION_TRIGGER | LLM09: Misinformation | 30 |
| **Total** | | **150** |

**By OWASP risk:** LLM01 (60), LLM06 (30), LLM09 (60).
**By severity:** CRITICAL (38), HIGH (92), MEDIUM (20).
**By source:** curated (50), mutation-engine (100).

## Loading

```python
from datasets import load_dataset

ds = load_dataset("9mark9/llm-redteam-owasp-prompts", split="train")
print(len(ds))                       # 150
print(ds[0])                         # first payload record

# Target one attack category for evaluation
jailbreaks = ds.filter(lambda r: r["category"] == "JAILBREAK")
```

See [`load_example.py`](./load_example.py) for a fuller example.

## Ethical Use & Responsible Disclosure

> **These payloads are for defensive evaluation only.**

By using this dataset you agree to apply it solely to **assess, harden, and
improve** the safety of LLM systems you are authorized to test. Do **not** use
it to attack third-party systems, to extract genuinely harmful instructions
from models, or to circumvent safety controls in production deployments. Many
prompts deliberately reference dangerous topics as *bait* to test whether a
model refuses — the dataset contains no actual harmful instructions, only the
adversarial prompts that should be refused. If you discover a safety regression
in a deployed model, follow the vendor's **responsible-disclosure** process
before publishing details.

## Source & Links

- **Source project (GitHub):** https://github.com/shaikn6/llm-safety-auditor
- **Live demo (HF Space):** [`9mark9/llm-safety-auditor`](https://huggingface.co/spaces/9mark9/llm-safety-auditor)
- **License:** MIT (inherited from the source project).
- **Contact:** nagizaazs@gmail.com

## Citation

```bibtex
@misc{llm_redteam_owasp_prompts,
  title  = {LLM Red-Team Prompts — OWASP LLM Top 10},
  author = {llm-safety-auditor contributors},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/9mark9/llm-redteam-owasp-prompts}}
}
```