break-ai-fix-loops
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data such as command outputs and evidence records, which presents an attack surface where malicious content in the project files could attempt to influence agent behavior.
- Ingestion points: The agent reads command results and JSON evidence records processed by the
scripts/fingerprint.pyscript. - Boundary markers: The skill defines a strict JSON schema for fingerprints and provides a specific template for the evidence ledger in
references/evidence-ledger.md. - Capability inventory: The agent has capabilities to modify the local filesystem and execute verification commands as part of the debugging process.
- Sanitization: The
scripts/fingerprint.pyscript performs validation to reject unknown fields and non-conforming data types. The instructions also mandate redacting credentials and sensitive data before recording evidence. - [COMMAND_EXECUTION]: The unit tests in
scripts/test_fingerprint.pyusesubprocess.runto verify the CLI behavior of the companion script. This execution is confined to testing the local python environment and the skill's own scripts using argument lists rather than shell strings.
Audit Metadata