harness-engineering
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The methodology outlines a system where automated agents ('GC Agents') scan repository artifacts, including logs, source code, and documentation, to perform maintenance tasks like removing dead code or flagging stale docs. This creates a surface where untrusted data residing in these files could theoretically be used to influence agent behavior.
- Ingestion points: Source code, task-specific logs, and repository documentation (SKILL.md section 3.1).
- Boundary markers: The skill suggests using 'AGENTS.md' as a ledger, but does not explicitly define delimiters for external or untrusted data processed during automated GC tasks.
- Capability inventory: Includes capabilities for automated file modifications via linting (
bin/lint --fix) and creating Pull Requests (bin/gc-pr). - Sanitization: Relies on binary pass/fail CI gates and human review of generated PRs as the primary validation mechanisms.
Audit Metadata