The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a local HTTP proxy server (cdp-proxy.mjs) that allows the AI to execute arbitrary JavaScript code within browser tabs via the /eval endpoint. It also provides endpoints for simulating hardware-level mouse events (/clickAt) and browser navigation.
[DATA_EXFILTRATION]: The skill is specifically designed to attach to the user's primary browser instance ("daily Chrome") to leverage existing authenticated sessions. This provides the agent with access to sensitive user data, private accounts, and logged-in web services without requiring new credentials.
[DATA_EXFILTRATION]: The proxy includes a /setFiles capability that allows the agent to programmatically select local files from the user's filesystem and upload them to web forms, creating a direct path for the exfiltration of local data to external websites.
[COMMAND_EXECUTION]: The check-deps.mjs script automatically starts the CDP proxy as a detached background process. It also includes logic to intercept and block network requests from websites attempting to detect the local Chrome debugging port, employing stealth techniques to bypass anti-automation security measures.
[EXTERNAL_DOWNLOADS]: The instructions direct the agent to utilize r.jina.ai for content extraction, which involves transmitting user-supplied URLs and their contents to an external third-party service.
[PROMPT_INJECTION]: The skill exhibits a significant indirect prompt injection surface as it is designed to ingest and interpret uncontrolled data from various web platforms. (Ingestion points: WebSearch, WebFetch, curl, and CDP proxy; Boundary markers: None specified; Capability inventory: Browser automation, file upload, and file writing; Sanitization: None).

web-access