The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/quick_validate_skill.py utilizes subprocess.run() to execute the local test suite tests/test_skill_contract.py. This is a benign implementation used solely for internal validation of the skill's components and does not involve untrusted input.
[PROMPT_INJECTION]: The skill incorporates dedicated testing fixtures (tests/fixtures/adversarial.md) and instructions to handle 'overclaim pressure' from users. It is designed to refuse requests to ignore safety warnings or provide false confidence, demonstrating a strong security posture against direct prompt injection.
[SAFE]: The tests/test_skill_contract.py script uses the Python compile() function to perform syntax validation on the skill's own scripts. This operation is limited to generating a code object and does not execute the code, serving as a safe linting mechanism.
[SAFE]: The skill identifies ingestion surfaces for untrusted data (e.g., notebooks, SQL, schemas) as part of its primary function. It mitigates indirect prompt injection risks by requiring a structured output contract and prioritizing evidence-based findings over generic advice.

model-risk-compliance-memo