computer-use
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the user's screen and web pages, creating a surface for instructions embedded in that data to influence agent behavior.
- Ingestion points: The capture action (screenshots) and the browser_exec tool (DOM content) bring external data into the agent's context.
- Boundary markers: The skill provides explicit 'Safety' instructions to disregard instructions found in screenshots or web content, treating the user's original prompt as the sole source of truth.
- Capability inventory: The skill possesses powerful capabilities including simulated keyboard input (type), mouse control (click, drag), and interaction with the terminal and write_file tools.
- Sanitization: Relies on the agent's adherence to safety instructions rather than technical sanitization of visual or structural data.
- [EXTERNAL_DOWNLOADS]: The documentation references external resources for the underlying driver and provides installation instructions for secondary skill packs.
- Evidence: References to github.com/trycua/cua and instructions to run 'hermes computer-use install' and 'cua-driver skills install'.
- [COMMAND_EXECUTION]: The skill provides an interface for executing commands via terminal emulation and simulated keystrokes.
- Evidence: The type action and references to a terminal tool allow the agent to input text and execute shell commands, although a blocklist for dangerous patterns is mentioned.
Audit Metadata