glmv-caption
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection as it ingests and processes untrusted media files.
- Ingestion points: The skill ingests local images and remote URLs for images, videos, and documents through the
scripts/glmv_caption.pyscript. - Boundary markers: The skill constructs structured JSON payloads to separate media from the text prompt in the API request, although it does not include explicit system instructions for the model to ignore potential directives embedded within the processed media.
- Capability inventory: The Python script performs local file reading (specifically for image base64 encoding) and makes network POST requests to the official ZhiPu API endpoint (
open.bigmodel.cn). - Sanitization: The script includes robust validation for file extensions, image resolution limits, and file sizes to ensure compliance with API constraints, but it does not perform semantic sanitization of the content for adversarial instructions.
Audit Metadata