The Agent Skills Directory

[DYNAMIC_EXECUTION]: The file references/workflows.md provides a code snippet that demonstrates using pickle.load() to deserialize local model files (model.pkl). Python's pickle module is inherently unsafe for parsing untrusted data, creating a risk of arbitrary code execution if a malicious file is supplied, although the template includes a comment warning the user to only load trusted files.
[INDIRECT_PROMPT_INJECTION]: In SKILL.md, instructions direct the agent to fetch external data from arxiv.org URLs or APIs to update author and citation details dynamically at runtime. This external ingestion point functions as a tool output poisoning surface, allowing untrusted content from the external web resource to enter the agent's context without structural boundary markers or sanitization.

pymc