The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's primary execution mechanism involves generating temporary JavaScript files from agent-provided code and loading them dynamically using the require() function in run.js and connect-local.js.
[EXTERNAL_DOWNLOADS]: The runAccessibilityAudit function in lib/helpers.js downloads the axe-core accessibility engine from a well-known CDN (cdnjs.cloudflare.com) at runtime.
[COMMAND_EXECUTION]: The skill executes system commands for setup tasks (npm install, npx playwright install) and environment checks (pgrep, tasklist) using child_process.execSync.
[DATA_EXFILTRATION]: The skill provides capabilities to upload local files, screenshots, and videos to external services such as Vimeo and Bunny Stream using user-configured API keys.
[PROMPT_INJECTION]: Processing external web content via functions like getPageStructure introduces a risk of indirect prompt injection, where malicious instructions embedded in a webpage could influence agent behavior.
Ingestion points: External webpage content and ARIA snapshots processed in lib/helpers.js and lib/local-browser.js.
Boundary markers: None identified in the logic that interpolates or processes webpage text.
Capability inventory: Includes arbitrary script execution, network requests, and file system operations.
Sanitization: No specific sanitization or instruction-filtering is applied to external web content before it is presented to the agent.

browser