device-screen-verification

Pass

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes various system commands including adb for screen capture, sips for image processing, and starts a local Python HTTP server.
  • [DATA_EXFILTRATION]: The skill starts a web server (python3 -m http.server) inside the /tmp directory. While it explicitly binds to 127.0.0.1 to prevent network exposure, it exposes the entire contents of the /tmp directory to any other process or user on the local machine for the duration of the verification process.
  • [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted visual data from an Android device screen.
  • Ingestion points: Screen captures are taken from a connected Android device using adb exec-out screencap and saved to /tmp/device_screen.png in SKILL.md.
  • Boundary markers: No delimiters or ignore embedded instructions warnings are present in the vision prompt instructions.
  • Capability inventory: The agent possesses capabilities to execute shell commands (adb, sips, python), write local files, and navigate browser tools (browser_navigate).
  • Sanitization: No sanitization, validation, or filtering is performed on the visual content before it is processed by the vision model.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 12, 2026, 01:34 AM
Security Audit — agent-trust-hub — device-screen-verification