adversarial-review
Pass
Audited by Gen Agent Trust Hub on Jun 29, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The instructions require the agent to execute shell commands for testing, linting, and building, which are resolved from project files like README, AGENTS.md, or CI configurations. This represents the execution of project-defined logic as part of the verification process.
- [PROMPT_INJECTION]: The skill is designed to process untrusted data (source code and project documentation). It relies on a 'refute-by-default' cognitive framework rather than technical boundary markers to manage potential indirect prompt injections from the reviewed content. 1. Ingestion points: AGENTS.md, README, CI configurations, and source code diffs. 2. Boundary markers: The skill lacks technical delimiters for untrusted data but mandates a skeptical adversarial persona as a behavioral boundary. 3. Capability inventory: Shell command execution (git grep, project test/lint commands) and file system access. 4. Sanitization: No explicit sanitization or validation of the resolved commands is mentioned.
Audit Metadata