diagnosing-bugs
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process potentially untrusted external data, including network traces, log dumps, and captured artifacts. This creates an attack surface where malicious instructions embedded in these data sources could attempt to influence the agent's behavior.\n
- Ingestion points: Processing of HAR files, log dumps, and core dumps (SKILL.md, Phase 1).\n
- Boundary markers: No explicit delimiters or "ignore embedded instructions" warnings are required for the ingested data.\n
- Capability inventory: The agent is expected to execute shell commands, run tests, and perform HTTP requests (SKILL.md, Phase 1).\n
- Sanitization: The skill provides manual instructions for secret redaction but lacks automated sanitization or filtering of external content.\n- [DYNAMIC_EXECUTION]: The skill encourages the creation and execution of temporary scripts, throwaway harnesses, and test loops (Phase 1 and Phase 5). While these are generated from templates or defined by the agent during the debugging session, they involve runtime creation and execution of code.\n
- Evidence: Usage of
scripts/hitl-loop.template.shand instructions to spin up minimal system subsets or fuzz loops.
Audit Metadata