dual-verifier
Pass
Audited by Gen Agent Trust Hub on Aug 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes external content including source code, diffs, and test logs, creating an attack surface for instructions embedded in data.\n
- Ingestion points:
SKILL.mddefines process flows for reading repository files, branch diffs, and terminal outputs.\n - Boundary markers: The subagent prompt uses standard headers but lacks explicit delimiters or instructions to treat ingested code as data rather than instructions.\n
- Capability inventory: The skill allows filesystem access and command execution through subagents.\n
- Sanitization: No sanitization or escaping of the ingested code is described in the workflow.\n- [COMMAND_EXECUTION]: The skill instructs the agent to execute code from the repository being analyzed.\n
- Evidence: The prompt in
SKILL.mddirects the agent to 'Run them [tests] yourself if you can', which involves executing local files that may be influenced by the task being verified.
Audit Metadata