---
license: apache-2.0
language:
- en
size_categories:
- n<1K
task_categories:
- text-classification
tags:
- cybersecurity
- cloudtrail
- aws
- attack-reconstruction
---

# CybersecJSONSchemaBench — CloudTrail v6 (Leak-Free)

100 long-context CloudTrail attack-reconstruction tasks with synthetic MITRE ATT&CK chain injection into real benign flAWS logs.

**Public split:** prompt + answer_schema + records only. Gold answers and construction metadata are explicitly STRIPPED to prevent agent contamination. For scoring, see the companion gated dataset `achinta3/cybersec-v6-gold`.

**Stripped fields (vs the older `achinta3/cybersec-jsonschemabench-cloudtrail-v6`):**
- `record["source_file"]` — was a deterministic synthetic/benign oracle.
- Row-level `template_id`, `template_name`, `chain_event_count`, `phase_order`.
- `gold_answer`, `gold_metadata` — moved to the private gold dataset.
- `input_artifact.{slice_jsonl_path, path}`, `answer_schema_path` — path leakage.

**Provenance:** Built from the frozen v6.1.0 checkpoint (commit e84b3d6 of JsonSecurityBench).