pokeclaw-android-ai-agent
Warn
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: MEDIUMPRIVILEGE_ESCALATIONDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill requests
BIND_ACCESSIBILITY_SERVICEandBIND_NOTIFICATION_LISTENER_SERVICEpermissions. These are high-privilege Android permissions that allow the agent to intercept all screen content, monitor incoming notifications, and perform gestures on behalf of the user. - [DATA_EXFILTRATION]: The
CloudLLMEnginemodule allows the transmission of session history to external endpoints defined via environment variables (POKECLAW_CLOUD_ENDPOINT). Since the history contains content retrieved viagetScreenInfo(), this facilitates the transmission of sensitive information displayed on the screen to remote servers. - [COMMAND_EXECUTION]: Installation instructions include shell commands for building the project and installing the resulting APK onto a device (
./gradlew assembleDebug,adb install app/build/outputs/apk/debug/app-debug.apk). - [EXTERNAL_DOWNLOADS]: The
ModelManagerclass fetches large model files from Hugging Face's official repository (google/gemma-4-e2b-it-litert), which is a well-known service for hosting machine learning models. - [INDIRECT_PROMPT_INJECTION]: The agent processes untrusted data from the device screen via
getScreenInfo(). Malicious content displayed in web browsers, emails, or messaging apps could potentially influence the agent's behavior to perform unauthorized actions such as tapping buttons or sending messages.
Audit Metadata