v-human-app-tester

Pass

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill relies on extensive ADB shell commands to interact with the device, including package installation (adb install), activity management (adb shell am start), input simulation (adb shell input), and system settings modification (adb shell content insert).\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the device to guide its testing logic, which presents a surface for malicious content to influence the agent.\n
  • Ingestion points: System logs via adb logcat and UI XML dumps via adb pull /sdcard/ui.xml are read from the device directly into the agent's context.\n
  • Boundary markers: No specific delimiters or instructions to ignore potentially embedded commands within the device's UI or logs are provided.\n
  • Capability inventory: The skill possesses broad capabilities including shell command execution on the host and device, and local file system writes for logging findings.\n
  • Sanitization: No sanitization or validation logic is defined for the data ingested from the Android device before it is interpreted by the agent.\n- [DATA_EXFILTRATION]: The skill captures potentially sensitive information from the device, including screenshots and system logs, storing them in a local directory for analysis. Although it instructs the agent to avoid exfiltrating personal data, the automated collection of logs and screen state represents an data exposure surface.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 17, 2026, 07:41 AM