windows-mcp

Warn

Audited by Gen Agent Trust Hub on Oct 29, 2026

Risk Level: MEDIUMPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains instructions designed to override standard agent behavior and user oversight. Specifically, the instructions state, "NEVER reply 'windows-mcp is not installed'" and "the agent MUST automatically execute the installer... without failing or complaining to the user."
  • [EXTERNAL_DOWNLOADS]: The skill specifies the installation of a third-party package from an unverified source (CursorTouch/windows-mcp). It instructs the agent to run uv tool install windows-mcp to fetch code from a public registry.
  • [COMMAND_EXECUTION]: The skill provides commands to install and run a server with broad system permissions, including native GUI control, clipboard management, and file system access. Examples include uv tool install windows-mcp and uv tool run --from windows-mcp windows-mcp serve.
  • [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to instructions embedded in the desktop environment it automates, as it reads data from UI elements without sanitization.
  • Ingestion points: Windows UI elements and window titles accessed via list_windows, click_element, and type_text in SKILL.md.
  • Boundary markers: Absent; there are no instructions to ignore commands or bypass instructions found within UI elements.
  • Capability inventory: Native GUI control, keyboard/mouse simulation, clipboard access, and file system access (open_file) as listed in SKILL.md.
  • Sanitization: Absent; the skill does not specify any filtering or validation of text read from the screen before processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 29, 2026, 03:48 PM