full-self-improvement

Pass

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to autonomously modify the repository's instructions and configuration files based on the analysis of external human reactions found in previous interaction threads.
  • Ingestion points: Uses t3_thread_list and t3_thread_read (SKILL.md, Step 2 and 3) to fetch untrusted content from conversation history.
  • Boundary markers: There are no explicit delimiters or instructions to ignore embedded malicious prompts within the analyzed thread data.
  • Capability inventory: The skill has the authority to edit SKILL.md, references/*, and agents/openai.yaml files (SKILL.md, Scope and Step 5).
  • Sanitization: No sanitization or validation of the extracted human reactions is performed before using them to drive file edits.
  • Risk: A malicious user could provide corrections in a standard thread that are actually instructions intended to weaken security or introduce malicious behavior, which the skill would then apply to the codebase without further approval.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 7, 2026, 09:42 AM