The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied MDX files to determine diagram requirements, creating a vector where malicious text in the documentation could influence agent actions.
Ingestion points: Reads existing .mdx files (Page input) and concept descriptions (Source input) to design diagrams.
Boundary markers: The instructions do not define boundary markers or clear separators between the untrusted page content and the agent's logic, making it susceptible to embedded instructions.
Capability inventory: The agent is authorized to execute shell commands (npm, npx, node, curl), write new SVG files to the repository, and modify existing MDX source files.
Sanitization: There is no evidence of sanitization or filtering applied to the input text before it is used to generate file content or drive automation scripts.
[COMMAND_EXECUTION]: The workflow relies on executing several shell commands to manage dependencies, run a development server, and execute a local verification script.
Evidence: Commands include npm install, npm ci, npm run dev, and the execution of screenshot.mjs via Node.js.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to install a browser environment for visual testing, which involves downloading binaries from an external source.
Evidence: npx playwright install chromium fetches Chromium binaries from Microsoft's Playwright infrastructure.

docs-diagrams