videodb
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection vulnerability surface due to the ingestion and automated processing of untrusted external media.
- Ingestion points: Media files, public URLs, YouTube videos, and live RTSP/RTMP feeds are ingested via
coll.upload()andcoll.connect_rtstream()as documented inSKILL.mdandreference/rtstream-reference.md. - Boundary markers: Absent. The instructions for indexing visual or spoken content (e.g.,
video.index_scenes()orvideo.search()) do not encapsulate the content within structural delimiters or instruct the model to ignore malicious instructions embedded within the audio or visuals. - Capability inventory: The skill leverages the VideoDB Python SDK to execute server-side video compilation, timeline manipulation, text-to-speech generation, and automated content analysis. It also executes a local WebSocket event listener (
scripts/ws_listener.py) that writes event streams to disk. - Sanitization: No data validation or sanitization mechanisms are implemented to escape or clean generated transcript texts or visual scene descriptions before passing them to the collection-level text generation API (
coll.generate_text()).
Audit Metadata