The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The Android client (autojs-task-phone-client.js) connects to a PC relay server and downloads scripts via the /probe/ endpoint. These scripts are then executed on the mobile device using engines.execScriptFile, allowing the AI to run arbitrary JavaScript in a high-privilege environment (Accessibility Services).
[DATA_EXFILTRATION]: Multiple task templates enable the transfer of sensitive data from the phone to the PC. get_clipboard.js reads the device clipboard and uploads long text to the PC, while download_file.js allows the retrieval of any file accessible on the /sdcard/ partition.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect injection by processing untrusted data at multiple points:
Ingestion points: User-supplied strings are ingested through run_task.js and task templates like tap_text.js and input_text.js (file: scripts/run_task.js).
Boundary markers: The skill lack explicit delimiters or instructions to ignore embedded commands within user-supplied text strings.
Capability inventory: The mobile client has the capability to execute scripts (engines.execScriptFile), delete directories (files.removeDir), and perform network operations (http.postMultipart) (file: scripts/autojs-task-phone-client.js).
Sanitization: Path traversal protection is implemented in scripts/relay/utils/fsx.js to ensure files stay within expected directories.
[COMMAND_EXECUTION]: The AI agent is instructed to run local Node.js bridge scripts (such as run_task.js and deploy_project.js) to trigger remote actions. This allows the model to interact with the local shell to communicate with the relay server.
[DYNAMIC_EXECUTION]: The deploy_project.js script can dynamically bundle multiple local files into a ZIP archive and deploy them to the phone, followed by immediate execution of the project's entry point (main.js).

autojs-task-runner-yashu