ai-ml-data-science
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied JSON and JSONL specification files containing fields like feature names and descriptions, which are then interpolated into Markdown reports. This creates a surface for indirect prompt injection if the agent interprets the generated reports as instructions.
- Ingestion points: The scripts
ml_toolkit.py(via--inputflag) andleakage_scan.py(via--specflag) ingest data from local metadata files. - Boundary markers: The generated Markdown reports do not implement explicit boundary markers or security warnings to distinguish tool-generated structure from user-provided data.
- Capability inventory: The toolkit includes file-system read/write capabilities and a test runner script (
scripts/test_leakage_regressions.py) that utilizessubprocess.runto execute Python scripts. - Sanitization: User-provided strings from the input specifications are not sanitized or escaped before being placed into Markdown tables and descriptive sections.
- [COMMAND_EXECUTION]: The script
scripts/test_leakage_regressions.pyuses thesubprocessmodule to execute other scripts within the project for regression testing. - Evidence: The
invokefunction callssubprocess.run(args, ...)whereargsis constructed from the Python interpreter and local script paths. - Context: The command execution is used as a test runner for local validation of the toolkit's CLI functionality.
Audit Metadata