Consented Romantic Video
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided data to generate synthetic media content, creating a vulnerability surface for malicious instructions embedded in descriptions.
- Ingestion points: Natural language scene descriptions (setting, wardrobe, action, camera) and reference images defined in the 'Required inputs' section of SKILL.md.
- Boundary markers: The workflow mandates a 'Mandatory consent gate' and requires the agent to 'freeze the scene description' before execution to prevent adversarial prompt drift.
- Capability inventory: The skill is granted 'media.generate_video', 'media.upload_file', and 'workspace-write' capabilities in its configuration frontmatter.
- Sanitization: The 'Execute and review' section requires a multi-faceted manual check for identity boundaries, unwanted nudity, and synthetic-media disclosure labels.
Audit Metadata