reviewing-code
Pass
Audited by Gen Agent Trust Hub on Apr 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local shell commands including
git diffto identify code changes and project-specific lint or test commands to verify fixes. These operations are essential for the skill's functionality and are confined to the local repository environment. - [PROMPT_INJECTION]: The skill processes untrusted external code which presents an indirect prompt injection surface.
- Ingestion points: Untrusted code is ingested via
git diffoutput and file reads in the scoping and fixing phases (SKILL.md). - Boundary markers: No explicit delimiters or specific 'ignore' instructions are used to wrap the untrusted code content provided to the reviewer subagents.
- Capability inventory: The agent has the capability to modify files on the local disk (Phase 4) and execute arbitrary shell commands defined by the project's testing configuration.
- Sanitization: No explicit sanitization of input code is performed; however, the skill's multi-agent loop, independent verification phases, and high confidence threshold (70+) serve as inherent structural safeguards against unintended execution.
Audit Metadata