skills/irpsv/ai-bro/bro-review-code/Gen Agent Trust Hub

bro-review-code

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external code, diffs, and pull requests. This untrusted data is interpolated into prompts for specialized subagents, creating a vulnerability where malicious instructions in code comments could influence agent behavior.
  • Ingestion points: The skill reads the repository history, diffs, and PR content as described in SKILL.md and passes them to subagents.
  • Boundary markers: The subagent prompt templates (e.g., correctness-review-prompt.md, security-review-prompt.md) use simple placeholders like <review context> without explicit delimiters or instructions to ignore embedded commands in the analyzed code.
  • Capability inventory: The agent has tool-based read access to the local file system and repository history.
  • Sanitization: No input sanitization or escaping of the code data is implemented before it is processed by the LLM.
  • [METADATA_POISONING]: The references/subagent-model-tiers.md file provides deceptive information by listing non-existent and speculative AI models (such as gpt-5.5, gpt-5.6-sol, grok-4.6, and claude-fable-5.1) as target execution environments for its subagents.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 07:20 AM
Security Audit — agent-trust-hub — bro-review-code