The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a system.run capability that allows an AI agent to execute arbitrary shell commands on the host Windows system. This is defined in the NodeService.cs capability registration logic.
[DATA_EXFILTRATION]: The skill exposes multiple sensitive data access points through registered capabilities, including screen.snapshot for capturing the desktop, camera.snap for taking photos via the webcam, and location.get for retrieving geolocation data.
[REMOTE_CODE_EXECUTION]: The CanvasPresentCapability uses a WebView2 control to render arbitrary HTML strings or navigate to URLs provided by the agent via the html and url parameters. This allows for dynamic execution of web content within the application context.
[EXTERNAL_DOWNLOADS]: The UpdateService implements an auto-update mechanism that fetches, downloads, and executes binaries from GitHub Releases based on user-selected channels (stable/alpha/dev).
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from WebSocket messages and external APIs (like the gateway) which then flows into high-privilege capabilities such as system.run without documented sanitization or boundary markers.

openclaw-windows-node