The Agent Skills Directory

[DATA_EXFILTRATION]: The understand_image tool is designed to accept local file paths, including absolute paths (e.g., /Users/username/Documents/image.png). The skill's documentation explicitly describes reading these files, converting their contents to base64, and sending the resulting data to an external API endpoint ({api_host}/v1/coding_plan/vlm). This capability allows the agent to access and exfiltrate arbitrary files from the user's system if directed to do so.
[INDIRECT_PROMPT_INJECTION]: The skill enables the ingestion of untrusted external data through image analysis and OCR (Optical Character Recognition). Malicious instructions embedded within images could be extracted and processed by the agent, potentially influencing its subsequent behavior.
Ingestion points: The understand_image tool in SKILL.md accepts external URLs and local files.
Boundary markers: None provided in the tool instructions to delimit or ignore instructions found within images.
Capability inventory: The skill has the ability to read local files and perform network POST requests to an external API as documented in SKILL.md.
Sanitization: There is no evidence of filtering or sanitizing the text extracted from images before it is returned to the agent's context.

minimax-image-understanding