adversarial-reviewer-acrazie
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is inherently designed to ingest and analyze untrusted external data in the form of code diffs and pull request content.
- Ingestion points: The verification procedure in SKILL.md (Phase 1) specifies the extraction of raw unified diffs, PR patches, and staged changes.
- Boundary markers: The instructions contain explicit directives to "Discard author commentary, PR summaries, and commit messages" and "Strip away the author's narrative excuses... to eliminate confirmation bias," which serves as a prompt-level boundary to mitigate the influence of embedded instructions.
- Capability inventory: The skill and its associated scripts lack dangerous capabilities such as network access, system modification, or arbitrary command execution. The provided Python test script (
scripts/test_adaptive_catalogue_contract.py) is restricted to read-only operations within the local file system. - Sanitization: No explicit sanitization or escaping of the ingested code diffs is performed, which is standard for an analysis-only skill.
Audit Metadata