The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill mandates the automatic installation and execution of third-party packages at runtime using package managers like npx and uv without version pinning or integrity verification. This allows for the execution of arbitrary remote code on the host system. Evidence: Instructions in SKILL.md to execute 'npx -y chrome-devtools-mcp@latest', 'uv tool run windows-mcp serve', and 'uvx devstudio-mcp' on-the-fly.
[EXTERNAL_DOWNLOADS]: The skill fetches multiple external dependencies from public registries (NPM and PyPI) and external repositories. Evidence: Downloads for 'chrome-devtools-mcp', 'windows-mcp', 'devstudio-mcp', 'web-infra-dev/midscene-skills', and 'am-will/codex-skills' are defined in the 3-Tier Architecture and fallback sections.
[COMMAND_EXECUTION]: The agent is prompted to execute shell commands to manage its own automation engines and interact with the operating system. Evidence: Commands such as 'uv tool run windows-mcp serve' and 'npx skills add' are provided for direct execution by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill has a high attack surface for indirect injection as it is designed to ingest and act upon data from external, untrusted sources such as web browsers and desktop accessibility trees. Evidence: 1. Ingestion points: DOM tree traversal, CSS extraction, and Windows UIA tree inspection (SKILL.md). 2. Boundary markers: Absent. There are no instructions defining delimiters or warnings to ignore instructions embedded in the UI. 3. Capability inventory: High capabilities including mouse/keyboard control, active window switching, and native OS file dialog interaction (SKILL.md). 4. Sanitization: Absent. The skill does not mention filtering or validating UI text before it is processed by the agent's logic.

computer-use