The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted local MATLAB (.m) and data (.mat) files via triage tools like scan_m_code.py and inventory_mat_file.py.
Ingestion points: Tools read local filesystem content provided by the user (e.g., scripts/scan_m_code.py, scripts/inventory_mat_file.py).
Boundary markers: The instructions explicitly warn the agent: "Never run an untrusted .m, .mlx, MEX binary, MAT file... Static review does not prove safety." It emphasizes treat all external files as untrusted.
Capability inventory: The skill includes tools to read files, calculate hashes (reproducibility_report.py), and write new scaffold files (generate_function_scaffold.py). It does not include tools for arbitrary shell execution or network exfiltration.
Sanitization: The _common.py library implements robust path validation (rejecting symlinks, traversal, and URL-like strings) and identifier validation to prevent injection into generated code or command-line plans.
[SAFE]: The skill references several trusted organizations and well-known services (MathWorks, GNU, arXiv, SciPy, H5PY) for documentation, package compatibility, and citation purposes. These references are documented neutrally as sources of truth for the skill's primary function.

matlab