adversarial-qc
Pass
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes standard system utilities (Chromium) to generate PDF certificates from HTML templates. The implementation is security-conscious, explicitly instructing the agent to maintain the renderer sandbox and warning against the use of insecure flags like
--no-sandbox. - [PROMPT_INJECTION]: The skill processes untrusted 'deliverable' text, which is an inherent surface for indirect prompt injection. This is mitigated by its adversarial architecture, where two agents independently verify the content against checklists. Disagreements or unverified claims are flagged for human review, and agents are instructed to be 'skeptical' and 'evidence-based'.
- [SAFE]: The skill includes comprehensive documentation regarding its scope, legal use, and professional responsibility. It provides clear escalation triggers that require human sign-off for sensitive deliverables (e.g., legal filings), regardless of the automated QC result.
- [DATA_EXFILTRATION]: No unauthorized network operations were found. File operations are pinned to local matter workspaces or designated output directories, ensuring data remains within the user's controlled environment.
Audit Metadata