The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a potential surface for indirect prompt injection due to its primary function of analyzing user-provided manuscripts.
Ingestion points: User-provided manuscripts (PDF, Word, or Markdown) and auxiliary CSV data files are read into the agent's context during the intake and numerical verification phases.
Boundary markers: The instructions do not consistently implement explicit delimiters or markers to isolate untrusted manuscript content from the skill's logical flow.
Capability inventory: The skill uses the Bash, Read, Write, and Edit tools to perform its checks, including running multiple local Python scripts.
Sanitization: The skill largely mitigates injection risks by delegating complex analysis (arithmetic, style, and consistency checks) to deterministic Python scripts using regular expressions, rather than having the LLM directly reason over unverified text.
[DYNAMIC_EXECUTION]: A testing utility in scripts/confounding_findings_challenge/verify.sh utilizes a Python one-liner with eval() to parse and extract values from JSON artifacts during regression testing. While eval() is a dynamic execution sink, its use in this context is confined to local tests with developer-controlled input strings.

self-review