The Agent Skills Directory

[COMMAND_EXECUTION]: The driver.mjs script accepts an --eval argument which is executed within the Electron renderer process using eval(). This provides a mechanism for arbitrary JavaScript execution inside the application context. The script also explicitly disables the Electron sandbox (ELECTRON_DISABLE_SANDBOX: '1'), which increases the potential impact of any code executed within the renderer.\n- [DATA_EXFILTRATION]: The skill can extract and return information from the application UI via the --testid flag, the results of the --eval expression, and the window.title() call. This allows sensitive data displayed in the app to be read into the agent's context.\n- [DATA_EXFILTRATION]: The script allows specifying an output path for screenshots via the --out flag. Because this path is resolved using resolve(process.cwd(), ...), it could be used to overwrite files outside the intended directory if a specific path is provided.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes and returns data from the application's UI, which may contain untrusted content.\n
Ingestion points: UI text fetched via --testid, the application title, and results from evaluated JavaScript expressions in driver.mjs.\n
Boundary markers: None present; the skill does not instruct the agent to distinguish between UI data and instructions.\n
Capability inventory: The skill can run shell commands (building via bun, executing via node) and write files to the disk.\n
Sanitization: None; data read from the application is returned directly to the agent's output stream.

run-desktop