visual-image-understanding
Pass
Audited by Gen Agent Trust Hub on Aug 27, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructions in
SKILL.mdinclude a command to installuv(a fast Python package manager) fromastral.shif it is missing. This is a common developer tool from a recognized provider (Astral). The script also automatically manages thePillowdependency viauv's inline script metadata. - [COMMAND_EXECUTION]: The skill operates by executing a Python script (
visual_runtime.py) usinguv run. This script performs local image operations (resizing, encoding) and sends requests to an external API. - [DATA_EXFILTRATION]: The skill reads local image files, converts them to Base64, and sends them to a remote vision model API endpoint. The endpoint and model are configurable via user-provided environment variables (
VISUAL_BASE_URL,VISUAL_MODEL,VISUAL_API_KEY), which is standard for BYO-Key (Bring Your Own Key) utility skills. - [INDIRECT_PROMPT_INJECTION]: The skill processes image content which could contain text-based instructions. It acts as an image-to-text bridge. While this presents an attack surface where a malicious image could contain instructions for the agent, the risk is mitigated by the fact that the output is intended to be a descriptive Markdown summary for the user, and the script does not possess capabilities to perform dangerous system actions based on that text.
Audit Metadata