soia-dev-review-panel

Pass

Audited by Gen Agent Trust Hub on Aug 6, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill ingests and analyzes external, untrusted content including code diffs and local directory files, which constitutes a surface for indirect prompt injection.
  • Ingestion points: Step 0 of SKILL.md defines diff text and directory paths (skills/<name>/) as review targets.
  • Boundary markers: Step 4 instructions require the agent to cross-verify findings against original source materials with a "refute-by-default" posture.
  • Capability inventory: The skill is strictly read-only and does not perform edits or merges; it relies on command-line tools for syntax verification only.
  • Sanitization: No specific data sanitization is mentioned, relying instead on cognitive partitioning review steps.
  • [COMMAND_EXECUTION]: SKILL.md (Step 1, Command Correctness lens) instructs the agent to use CLI tools like gh to verify command syntax using non-destructive, query-only flags such as --help or --dry-run.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 6, 2026, 08:00 AM
Security Audit — agent-trust-hub — soia-dev-review-panel