diagnosing-bugs-with-docs
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data such as system logs, network traces, and captured artifacts to diagnose issues, which constitutes an indirect prompt injection surface.
- Ingestion points: Reads system logs, network traces, and user-supplied artifacts during Phase 1 and Phase 2 in SKILL.md.
- Boundary markers: The skill lacks explicit boundary markers for data processing, though it provides methodology to separate facts from inference.
- Capability inventory: The skill uses shell execution and network operations (curl/HTTP) to build feedback loops as described in SKILL.md.
- Sanitization: Instructs the agent to redact secrets and authentication headers from outputs before displaying them.
- [COMMAND_EXECUTION]: The skill's primary function involves creating and running reproduction scripts (feedback loops) which include shell commands and network requests.
- Evidence: Phase 1 of SKILL.md encourages the use of curl, CLI invocations, and custom bash scripts (scripts/hitl-loop.template.sh) to reproduce bugs.
- [DATA_EXFILTRATION]: The skill accesses local diagnostic vaults and project-specific knowledge bases to inform its diagnosis.
- Evidence: Instructions in Phase 0 and Phase 6 of SKILL.md reference reading and writing to ~/ops and project-specific entity alias tables.
Audit Metadata