ad-design

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from websites and brand documents. It mitigates risk by instructing the agent to treat all retrieved content as data, never instructions, and specifically requires the agent to report any detected injection attempts as findings rather than obeying them.\n- [DATA_EXFILTRATION]: The skill interacts with ad account libraries for asset staging. It maintains a strict 'human-in-the-loop' policy, ensuring no files are uploaded until a user manually selects them. It also adheres to security rules that forbid handling or persisting credentials.\n- [PROMPT_INJECTION]: Static analysis flagged a quality-control instruction ('Do not return a draft with a warning on it') as concealment. However, context shows this is a safety measure intended to prevent users from accidentally deploying incomplete or non-compliant ad copy, as users often ignore warnings when copying output.\n- [SAFE]: The skill references a comprehensive security document (references/agent-security.md) that mandates HTML escaping for interpolated content, forbids the use of dangerous HTML attributes, and enforces strict data privacy rules regarding protected categories.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:23 PM