research-loop
Pass
Audited by Gen Agent Trust Hub on Aug 12, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill handles potentially untrusted data from the repository codebase and evaluation harnesses.
- Ingestion points: implementation files, evaluation data, and harness scripts in the repository workspace.
- Boundary markers: usage of isolated run directories (
.research-loop/), separate Git worktrees, and recommended read-only mounts for evaluators. - Capability inventory: Git file operations, candidate modification, and shell command execution for evaluation purposes.
- Sanitization: mandatory secret redaction from output logs and SHA-256 fingerprinting of immutable files to prevent environmental tampering.
- [DYNAMIC_EXECUTION]: The skill automates the execution of evaluation commands defined in the repository.
- Evidence: instructions to "Run evaluation from a trusted boundary" and execute candidate logic to extract metrics.
- Mitigation: the protocol strictly requires a trusted environment (e.g., containerization) and integrity verification before and after execution to prevent candidate code from compromising the host.
Audit Metadata