The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The SyntaxVerifier and IVFEngine implementations in SKILL.md use the Python compile() function in "exec" mode and subprocess.run with node -e on strings generated by the AI agent. This pattern executes untrusted code during the verification phase, which could be exploited if an attacker influences the agent's output.
[REMOTE_CODE_EXECUTION]: SKILL.md contains a LintVerifier that executes npx eslint on generated files. Using npx without version pinning or a lockfile allows the environment to download and run arbitrary code from the public npm registry at runtime.
[DATA_EXFILTRATION]: The AsyncApprovalWorkflow in references/hitl-checkpoint-design.md sends POST requests to a user-configurable webhook_url containing artifact summaries and session context. While designed for human approval workflows, this creates a potential path for data exfiltration to unauthorized endpoints.
[SAFE]: Multiple reference documents (e.g., references/architecture-patterns.md, references/performance-optimization.md) contain 150 identical sections repeating the same technical paragraphs. This extreme redundancy acts as significant noise to complicate security auditing, although no hidden payloads were detected in the repetition patterns.

feedback-loops