lov-describe-image
Pass
Audited by Gen Agent Trust Hub on Aug 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script uses subprocess.run to call the macOS sips utility for image resizing. This is a controlled execution with no shell involvement, using a predefined argument list to prevent shell injection, and is central to the skill's stated purpose for handling large images.
- [EXTERNAL_DOWNLOADS]: The skill transmits image data and user-provided questions to the Zhipu AI API at open.bigmodel.cn. This is the core functionality of the skill and utilizes an established service provider for vision processing.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) as it processes untrusted image content and user questions.
- Ingestion points: Reads image file data from the local filesystem and user-supplied strings via script arguments in scripts/describe_image.py.
- Boundary markers: None implemented; the vision model's output is returned directly to the calling agent without delimiters to separate it from other instructions.
- Capability inventory: The script has permissions to read local files (restricted by extension), execute the sips system command, and make network requests to an external API.
- Sanitization: The skill validates input by file extension (.png, .jpg, etc.), but it does not sanitize or filter the text output generated by the vision model before it enters the agent's context.
Audit Metadata