The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions direct the user to clone the framework repository from a third-party source: https://github.com/zai-org/Open-AutoGLM.git.
[COMMAND_EXECUTION]: The skill uses Python's subprocess module to execute adb (Android Debug Bridge) and hdc (HarmonyOS Device Connector) commands. These commands control device hardware, manage applications, and simulate user input (taps, swipes, key events).
[DATA_EXFILTRATION]: The recommended vLLM model server deployment configuration includes the flag --allowed-local-media-path /. This configuration allows the model service to access any file on the host machine's root filesystem, presenting a high risk of sensitive data exposure if the model server is compromised or receives malicious prompts.
[INDIRECT_PROMPT_INJECTION]: The agent operates by processing screenshots of mobile device screens. This creates a surface for indirect prompt injection, where malicious content displayed within a third-party app or web page could influence the vision-language model to perform unauthorized actions on the device.
Ingestion points: Screenshots captured via adb exec-out screencap -p (in SKILL.md).
Boundary markers: The model uses structured output tags (<think> and <answer>), but these are not sufficient to prevent adversarial content within the processed images.
Capability inventory: Full device control via ADB/HDC, including application launching, text input, and gesture simulation.
Sanitization: The skill mentions a confirm_sensitive=True parameter to prompt users before high-risk actions like purchases, which serves as a mitigation strategy.
[CREDENTIALS_UNSAFE]: The documentation includes instructions for setting sensitive API keys (BIGMODEL_API_KEY, MODELSCOPE_API_KEY) via environment variables. While placeholders are used, the automation of these tasks often leads to secrets being stored in shell history or insecure configuration files.

open-autoglm-phone-agent