The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted evaluation datasets to measure agent performance and safety. Ingestion points: .jsonl data files processed by scripts such as goldenset-audit.py and judge-calibration.py. Boundary markers: The skill provides templates like assets/judge-rubric.template.md that utilize explicit delimiters (<<< >>>) and instructions to focus on specific criteria. Capability inventory: Standard Python 3 scripts performing file I/O and shell output; no network exfiltration or dangerous subprocess calls are present. Sanitization: The scripts use standard JSON parsing and perform rigorous statistical validation to interpret data. This surface is intentional and documented as part of the skill's core purpose for evaluation.
[SAFE]: All included scripts (eval-baseline.py, goldenset-audit.py, judge-calibration.py) are implemented using standard Python libraries and operate entirely offline. No hardcoded credentials or access to sensitive local environment paths were detected.
[SAFE]: The presence of prompt injection patterns in assets/golden-set.example.jsonl (e.g., 'Ignore your refund policy') is a safe finding, as these are clearly labeled as adversarial test cases designed to verify the safety and refusal capabilities of the agent under test.

evals-ops