panel-review-loop
Pass
Audited by Gen Agent Trust Hub on Jun 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to autonomously modify files in the working tree and execute the
panel-reviewtoolchain (which involves shell script execution) without intermediate user approval. - [EXTERNAL_DOWNLOADS]: The documentation describes installing the skill and its dependencies from the vendor's repository (
catena-labs/dev-skills) using thenpxutility. - [PROMPT_INJECTION]: The instructions explicitly direct the agent to operate with high autonomy, stating it should "never pause for your input until it's done" and avoid "per-round check-ins." This instruction overrides typical interactive safety checks and human-in-the-loop verification processes.
- [COMMAND_EXECUTION]: The skill processes findings from external agents (panelists) and uses them to generate and apply code fixes. This creates a surface for indirect prompt injection where malicious content in the codebase being reviewed could influence the suggestions, leading the agent to autonomously write insecure code to the local filesystem.
Audit Metadata