chatgpt-image-ad
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user prompts and reference images, which are then used by the agent to interact with an external API and write files to the local system.\n
- Ingestion points: User-provided seed prompts and reference image paths collected during Phase 2 of the workflow.\n
- Boundary markers: The skill implements specific safety suffixes (NO_CHROME_SUFFIX, SAFE_ZONE_SUFFIX, GLYPH_SAFETY_SUFFIX) and a dedicated prompt rewriting phase (Phase 3) to constrain model behavior.\n
- Capability inventory: The
scripts/generate_image.pyscript performs network operations (POST/PUT/GET) to the Arcads API and writes generated image files to the local filesystem.\n - Sanitization: A mandatory Visual QA phase (Phase 6) is included where the agent inspects the output for defects and performs retries if necessary, providing a verification step for the generated content.
Audit Metadata