review-fix

Pass

Audited by Gen Agent Trust Hub on Sep 23, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data to drive its core logic, creating an attack surface for indirect prompt injection.
  • Ingestion points: Data is ingested from user-provided briefs, Git PR/MR metadata, issue descriptions, and repository-specific documentation in SKILL.md (Establish the scope contract).
  • Boundary markers: There are no explicit delimiters or sanitization routines defined to isolate external content from the agent's instructions.
  • Capability inventory: The skill possesses significant capabilities, including the ability to apply code edits (repairs) and execute shell commands for testing and linting (Verification Gate) as described in SKILL.md.
  • Sanitization: The skill lacks explicit sanitization or filtering of the content retrieved from external forges or PR bodies.
  • [COMMAND_EXECUTION]: To verify repairs, the skill executes local scripts such as tests, linters, and type checkers. If a repository contains malicious configuration files (e.g., a poisoned package.json test script or pytest config), this execution path could be exploited to run arbitrary code on the local system.
  • [DYNAMIC_EXECUTION]: The orchestrator dynamically identifies and loads other 'Review Skills' based on the user's input brief. Although it includes a preflight check to ensure these skills are report-only and do not perform mutations, the dynamic resolution of these components based on untrusted text is a potential risk vector.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 23, 2026, 07:08 AM
Security Audit — agent-trust-hub — review-fix