investigate
Pass
Audited by Gen Agent Trust Hub on Mar 25, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and acting upon untrusted data from external and local sources without explicit boundaries or sanitization.
- Ingestion points: The agent reads arbitrary codebase files, git logs, and external data via
WebSearchand documentation queries (SKILL.md, Steps 2 & 3). - Boundary markers: Absent. The instructions do not define delimiters or provide specific instructions to treat researched content as data rather than instructions.
- Capability inventory: The agent can modify files, execute tests, and invoke sub-agents (SKILL.md, Steps 2 & 5).
- Sanitization: No sanitization or validation of the gathered data is mentioned before it influences the proposed or applied code changes.
- [COMMAND_EXECUTION]: The skill uses commands to investigate the environment and verify fixes.
- Evidence: It executes
git logto inspect history and runs "relevant tests" to verify proposed solutions (SKILL.md, Steps 2 & 5).
Audit Metadata