The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The skill has a high-severity vulnerability surface because it ingests untrusted data from external sources and possesses extensive system capabilities.
Ingestion points: Processes local images and PDFs, as well as remote content via URL fetching (SKILL.md).
Boundary markers: Absent. The code samples demonstrate direct interpolation of data into the model context without delimiters or safety framing.
Capability inventory: Includes Bash, Read, Write, Edit, Glob, Grep, Task, and WebFetch.
Sanitization: No evidence of sanitization or content filtering for OCR text or visual data.
[Command Execution] (MEDIUM): The skill explicitly allows the Bash tool. While standard for some tasks, its presence significantly escalates the risk of RCE if the agent is manipulated by instructions embedded in a processed image or document.

vision-multimodal