interrogate
Pass
Audited by Gen Agent Trust Hub on Aug 9, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill instructs subagents to adopt an adversarial persona in 'references/reviewer-prompt.md' (e.g., 'You are an adversarial code reviewer', 'You are not here to be helpful', 'You're an adversary, not a cheerleader'). While functional for code review, these instructions utilize role-play and behavioral override patterns similar to those found in prompt injection techniques.\n- [PROMPT_INJECTION]: The skill's architecture is susceptible to indirect prompt injection.\n
- Ingestion points: In 'SKILL.md' (Step 1), the agent gathers context from git diffs and workspace files, which are untrusted external inputs.\n
- Boundary markers: The 'references/reviewer-prompt.md' template incorporates the untrusted content into a block but lacks explicit directives to the subagent to disregard instructions or commands that might be contained within the code under review.\n
- Capability inventory: The skill uses the 'parallel' tool to launch subagents and pass them the collected code for processing.\n
- Sanitization: The skill does not perform any sanitization or escaping of the code content before passing it to the subagents.\n- [COMMAND_EXECUTION]: The skill executes 'git diff' in 'SKILL.md' (Step 1) to identify changes for review. This command execution is appropriate and necessary for the skill's intended functionality.
Audit Metadata