The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/quick_validate_skill.py uses subprocess.run to execute internal contract tests. Additionally, tests/test_skill_contract.py uses the compile() function to validate the syntax of Python files within the skill folder. These are legitimate development and validation tools.
[PROMPT_INJECTION]: The skill contains specific instructions and test fixtures (e.g., tests/fixtures/adversarial.md) to train the agent to resist adversarial pressure and refuse to ignore safety warnings. However, because the skill processes untrusted input such as pipeline code and notebooks, it maintains a surface for indirect prompt injection.
[EXTERNAL_DOWNLOADS]: The documentation and evidence map reference various research papers and industry reports from established organizations like Microsoft, Anaconda, and academic publishers. These references are used for grounding the skill's logic and are considered safe.

leakage-adversary