adjudicate-regression
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by processing untrusted external data.
- Ingestion points: The skill ingests 'candidate', 'frozen sources', 'reproduction/replay packages', and 'logs' as specified in the 'When to Use' and 'Decide' sections of SKILL.md.
- Boundary markers: There are no defined delimiters or instructions to treat ingested data as untrusted, which could allow malicious instructions within logs or source code to influence the agent's adjudication.
- Capability inventory: Capabilities are limited to reading local documentation/evidence files and writing a 'decision.json' file. No network operations, shell execution, or dynamic code evaluation were detected.
- Sanitization: No sanitization or validation logic is defined for the content of the ingested files or logs.
Audit Metadata