The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill guides users to download the Droidrun Portal APK from the vendor's official GitHub repository to establish a connection between their physical device and the service.
[CREDENTIALS_UNSAFE]: The setup workflow instructs the agent to reveal the user's active API key (dr_sk_...) in the chat interface so the user can manually paste it into the mobile application. This exposure of sensitive credentials in the conversation logs represents a security risk.
[DATA_EXFILTRATION]: By design, the skill captures and transmits potentially sensitive information, including device screenshots and full UI accessibility trees, to the vendor's API endpoints (api.mobilerun.ai) to allow the agent to perceive and interact with the device.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes untrusted data found on the mobile device's screen.
Ingestion points: Screen content is ingested through the GET /devices/{deviceId}/ui-state and GET /devices/{deviceId}/screenshot endpoints.
Boundary markers: There are no explicit boundary markers or instructions to ignore embedded commands within the ingested screen data.
Capability inventory: The skill has extensive capabilities including simulating taps (POST /tap), typing text (POST /keyboard), swiping (POST /swipe), and executing autonomous multi-step tasks (POST /tasks).
Sanitization: There is no evidence of sanitization or filtering of the ingested screen UI elements before they are interpreted by the agent.

mobilerun