The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains logic to propose modifications to other skill files and code assets. While it can generate Python code (.py), it includes explicit security rules prohibiting the use of dangerous functions like eval(), exec(), or subprocess with user-supplied input. It also mandates that all changes pass existing validation scripts (validate_skills.sh and test_skills_assets.sh).
[PROMPT_INJECTION]: The skill incorporates an 'Indirect Prompt Injection' surface as it processes untrusted user feedback to create persistent instructions. However, this is mitigated by a mandatory human-in-the-loop approval process for every proposal, a strict 'never self-modify' rule for its own security instructions, and a requirement to verify that learnings originate from genuine problem-solving rather than prompt-based instruction echoing.

cuopt-skill-evolution