The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill requires the user to grant extensive system-level permissions, specifically Screen Recording and Accessibility, to the runtime environment (Terminal, VS Code, Node.js, and Python). These permissions allow the agent to monitor all screen activity and simulate user input across the entire OS.
Evidence: README.md provides detailed steps to enable these permissions and specifically instructs users on how to manually add system-level binaries (example: /usr/libexec/sshd-keygen-wrapper) to the Accessibility list using path-entry methods to bypass default application restrictions.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted visual data from the desktop environment, including web browsers and email clients, which introduces a high risk of indirect prompt injection.
Ingestion points: Screen capture and visual analysis of active application windows (Browsers, Mail, Finder).
Boundary markers: None. The visual "Brain" model processes the raw screen state without clear delimiters between user data and system instructions.
Capability inventory: The skill can perform UI actions (clicks, keyboard input), navigate websites, and execute shell scripts via the run_turix.sh bridge.
Sanitization: There is no evidence of sanitization or filtering for instructions embedded in the visual content processed by the agent.
[COMMAND_EXECUTION]: The skill relies on a helper script (scripts/run_turix.sh) to launch the external TuriX CUA agent.
Evidence: The script executes "$CONDA_PATH" run -n "$ENV_NAME" python examples/main.py which triggers the main execution loop of the visual agent.
[DYNAMIC_EXECUTION]: The helper script employs dynamic code generation to manage configuration states.
Evidence: scripts/run_turix.sh uses a Python heredoc (python3 << 'PYEOF') to dynamically execute Python code that reads, modifies, and overwrites the config.json file at runtime based on environment variables.

turix-mac