image-director
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-provided images and text for vision analysis and editing tasks (documented in
references/vision-decomposer.mdandreferences/editing.md). This ingestion of external data is a core feature but constitutes an attack surface for indirect prompt injection. - Ingestion points: External visual assets and text descriptions supplied by the user during the art direction workflow.
- Boundary markers: The skill employs structured output formats (such as the 5-slot template in
references/gpt-image.mdand JSON for spatial logic inreferences/nano-banana.md) to maintain the integrity of the agent's instructions. - Capability inventory: The skill does not provide the agent with dangerous capabilities like arbitrary shell execution, file system modification, or network requests based on the processed data.
- Sanitization: The instructions include 'Anti-Slop' filters and literal text delimiters to ensure that user input does not bypass the intended prompt structure.
Audit Metadata