The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes user-supplied CSV or TSV data files (e.g., concentration-time profiles) for mathematical analysis. This ingestion of untrusted external data represents a potential surface for indirect prompt injection.
Ingestion points: The read_table function in scripts/_common.py is utilized by all analytical scripts (nca.py, fit_compartmental.py, exposure_response.py, etc.) to parse input data from the --input argument.
Boundary markers: The scripts do not implement specific LLM boundary markers or delimiters when presenting analysis results back to the agent.
Capability inventory: The skill is configured with Bash tool access and file modification permissions (Write, Edit).
Sanitization: All scripts enforce strict numerical parsing (using parse_float) and validate inputs against expected scientific ranges and enumerations (e.g., EVID event codes in check_popk_dataset.py), which effectively sanitizes the data against non-numeric instruction injection.
[EXTERNAL_DOWNLOADS]: The skill's documentation (references/software-ecosystem.md) references several legitimate scientific dependencies, including pharmpy-core, numpy, scipy, chi-drm, pints, and lmfit. These are well-known packages in the pharmacometrics and scientific computing communities and do not constitute a security risk. The documentation also correctly notes that the PKPy framework is only available via GitHub and not on PyPI, serving as a helpful technical reference for the user.

pkpd-modeling