footage-analysis
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill utilizes vendor-specific tools for media analysis, such as picsart_media_probe and picsart_media_transcribe, to perform its stated functions. No unauthorized third-party tool access or credential requests were detected.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from video and audio sources (talking footage). While this creates an attack surface, the instructions emphasize extracting structured JSON data with specific schema requirements, which acts as a boundary against embedded instructions.
- Ingestion points: User-provided video files processed via media description and transcription tools (SKILL.md).
- Boundary markers: Explicit instructions in references/talking-footage.md to return only one JSON object with predefined fields (moments, dead_ranges, people, shots, register).
- Capability inventory: Media probing (picsart_media_probe), visual description (picsart_media_describe_video), and audio transcription (picsart_media_transcribe) tools.
- Sanitization: The skill enforces a strict JSON schema for the output of analysis tools, which limits the potential for the model to execute instructions embedded within the analyzed media.
Audit Metadata