The Agent Skills Directory

[DYNAMIC_EXECUTION]: The file references/traceability-patterns.md provides an implementation example that utilizes the eval() function to process LaTeX formulas at compile time. While this code is presented as documentation, it encourages the use of unsafe dynamic execution patterns. If an agent implements this logic to evaluate untrusted formulas, it could lead to arbitrary code execution within the Python environment, despite the use of a restricted namespace.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to scan and process user-provided LaTeX source files and experiment code outputs to verify numeric consistency. This creates a surface where malicious instructions could be embedded within document content (e.g., within labels or values). If the verification script extracts these strings and the agent subsequently processes them as data, it could be influenced by embedded prompt injection payloads.
Ingestion points: The scripts/ref_numeric_values.py script reads from .tex files and code output logs (e.g., results.txt).
Boundary markers: There are no explicit markers or instructions to treat the content of the LaTeX files as untrusted or to ignore embedded instructions.
Capability inventory: The skill has the ability to read arbitrary local files through its verification script and write JSON reports to the filesystem.
Sanitization: The script extracts content via regular expressions but does not perform validation or sanitization on the resulting strings before they are included in the JSON report.

backward-traceability