volcengine-video-understanding
Pass
Audited by Gen Agent Trust Hub on Mar 2, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The
scripts/video_understand.pyscript accepts user-defined instructions and interpolates them directly into a JSON payload sent to the Volcengine API. This presents a surface for indirect prompt injection. - Ingestion points: The
instructionargument and the localvideo_pathprocessed inscripts/video_understand.py. - Boundary markers: No delimiters or specific system instructions are included to prevent the model from following potentially malicious instructions embedded in the input text or detected within the video content.
- Capability inventory: The script performs network operations using the
requestslibrary to transmit data to the remote API and process the returned analysis. - Sanitization: There is no evidence of input validation or escaping for the user-supplied text before it is sent to the Large Language Model endpoint.
- [EXTERNAL_DOWNLOADS]: The skill transmits video data and metadata to
https://ark.cn-beijing.volces.com. This is the official endpoint for Volcengine (ByteDance's cloud service), which is a well-known technology provider. - [COMMAND_EXECUTION]: The skill is designed to be executed via the command line to process files on the local filesystem using the provided Python script
scripts/video_understand.py.
Audit Metadata