The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill contains instructions for the agent to facilitate the installation of a driver by downloading and executing scripts from an external domain (cua.ai). Specifically, it uses high-risk patterns such as piping a remote script to a subshell (/bin/bash -c "$(curl ...)") on macOS/Linux and using iex (Invoke-Expression) with remote content on Windows. Although the skill explicitly directs the agent to ask the user for authorization before installation, the execution method itself is inherently risky.
[COMMAND_EXECUTION]: The skill relies on the execution of local CLI commands (cua-driver) to interact with the operating system. This allows the agent to perform a wide range of actions including window management, native menu invocation, and targeted input delivery, representing significant control over the user's environment.
[INDIRECT_PROMPT_INJECTION]: The agent is designed to ingest and act upon data from the user's desktop, such as accessibility trees and window snapshots. Because this data originates from third-party applications and web content, it could contain embedded instructions intended to manipulate the agent's logic. The skill mitigates this by instructing the agent to use semantic operations and verification loops, but the ingestion of untrusted UI data remains a vulnerability surface.
[DATA_EXFILTRATION]: The skill includes a tool for publishing desktop evidence (superset pages publish). While intended for documentation, the skill notes that desktop screenshots may inadvertently capture sensitive data like private messages or credentials. It relies on the agent's adherence to safety guidelines and the use of visibility flags to prevent unauthorized data exposure.

computer