The Agent Skills Directory

[COMMAND_EXECUTION]: The file references/workflows.md includes a code example demonstrating model persistence using pickle.dump() and pickle.load(). Specifically, the call saved = pickle.load(f) is inherently dangerous as it allows for arbitrary code execution if the .pkl file being loaded is malicious. While the documentation provides a textual warning to only trust your own files, the inclusion of this pattern in an automated agent's skill increases the risk of accidental or forced loading of untrusted objects.
[EXTERNAL_DOWNLOADS]: The skill instructions in SKILL.md require the installation of several Python packages, including pymc, arviz, and optional samplers like nutpie. While these are well-known libraries, the automated installation of packages from external registries is a sensitive operation that requires monitoring.
[PROMPT_INJECTION]: This skill identifies ingestion points for external data (predictors and outcomes) in assets/linear_regression_template.py and assets/hierarchical_model_template.py via variables X and y. The skill lacks explicit boundary markers or sanitization logic to prevent embedded instructions in processed data from influencing agent behavior. Given the skill's extensive capabilities (Bash, File Write), this provides a surface for indirect prompt injection attacks.
[SAFE]: The core logic within the diagnostic scripts (scripts/model_diagnostics.py) and model comparison utilities (scripts/model_comparison.py) follows standard statistical practices and does not exhibit malicious behavior.

pymc