The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download and run executable installers directly from an external GitHub repository.
Evidence: Invoke-WebRequest -Uri "https://github.com/openclaw/openclaw/releases/latest/download/OpenClawCompanion-Setup-x64.exe" -OutFile "OpenClawCompanion-Setup-x64.exe" and subsequent execution in SKILL.md.
[COMMAND_EXECUTION]: The skill enables a "Node Mode" that allows an AI agent to execute arbitrary system commands via PowerShell or other shells on the host machine.
Evidence: The system.run command capability described in the Node Mode configuration and code examples in SKILL.md.
[DATA_EXFILTRATION]: The skill grants the agent access to sensitive environmental data, including the ability to take screenshots and capture images from the camera, which can be sent back to the gateway.
Evidence: Documentation for screen.snapshot and camera.snap methods within the Node Mode section of SKILL.md.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests commands from an external WebSocket gateway to perform high-risk system operations.
Ingestion points: Messages received via the GatewayClient WebSocket connection processed in the EventHandler class in SKILL.md.
Boundary markers: The documentation describes a user-configurable allowCommands whitelist in ~/.openclaw/openclaw.json to restrict agent capabilities.
Capability inventory: Remote command execution (system.run), screen capture, camera access, and WebView2 script execution (canvas.eval).
Sanitization: No specific evidence of input sanitization or validation for the agent-supplied command arguments is provided in the documentation.

openclaw-windows-companion