The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external artifacts including training data, pipeline code, and computational notebooks as specified in SKILL.md and README.md.
Ingestion points: Training data, target definitions, split logic, feature lists, and pipeline code (referenced in SKILL.md under Required Inputs).
Boundary markers: The skill includes instructions to "Separate mechanical checks from expert judgment" and "Preserve raw data... unless the user explicitly asks for mutation," which serves as a delimiter for the agent's behavior.
Capability inventory: The skill possesses file-reading capabilities (scripts/split_leakage_probe.py), shell command execution (scripts/quick_validate_skill.py), and code compilation features (tests/test_skill_contract.py).
Sanitization: There is no explicit sanitization or filtering logic defined for the content of the analyzed data or code files.
[COMMAND_EXECUTION]: The validation script scripts/quick_validate_skill.py uses subprocess.run to execute the portable skill contract tests (tests/test_skill_contract.py). This is intended for local environment validation and portability checks but involves the execution of shell commands.
[DYNAMIC_EXECUTION]: The test suite tests/test_skill_contract.py utilizes the Python compile() function to verify that all .py files within the skill folder are syntactically valid and executable. While used strictly for testing purposes during development, this represents runtime code evaluation.

leakage-adversary