videodb
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external video and audio sources (files, URLs, live streams) which is then used to generate transcripts and visual indexes. This data is subsequently interpolated into LLM prompts for summarization and analysis.
- Ingestion points: Media is ingested via
coll.upload(url=...)as seen inSKILL.mdand transcripts are retrieved usingvideo.get_transcript_text()for use incoll.generate_text()inreference/generative.md. - Boundary markers: The provided examples interpolate transcript text directly into prompts (e.g.,
f"Summarize the key points discussed in this video:\n{transcript_text}") without using distinct boundary markers or explicit instructions to ignore potentially embedded commands. - Capability inventory: The skill utilizes network operations via the VideoDB SDK, executes shell commands through the Bash tool for recording and utility script management, and performs local file writes via the
ws_listener.pyscript. - Sanitization: No sanitization or validation of the extracted transcript or visual index text is performed before it is passed to the LLM.
Audit Metadata