agent-device

Pass

Audited by Gen Agent Trust Hub on Apr 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes system tools such as adb, simctl, sips, and ffmpeg for managing device state and processing media. It also suggests generating and executing local bash scripts to perform interaction sequences quickly.
  • [DATA_EXFILTRATION]: Provides capabilities to read the iOS clipboard and pull files from Android devices using adb pull. These operations are presented as standard features for mobile application testing and interaction.
  • [PROMPT_INJECTION]: The skill contains a surface for indirect prompt injection by ingesting untrusted data from the device UI. \n- Ingestion points: Application accessibility tree snapshots (labels, identifiers, values) via agent-device snapshot. \n- Boundary markers: None mentioned for identifying or isolating embedded instructions within UI elements. \n- Capability inventory: Subprocess calls including agent-device press, type, swipe, and adb shell. \n- Sanitization: No validation or escaping of element labels is described before being used to compute coordinates.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 13, 2026, 02:35 AM