harmonyos-device-automation

Pass

Audited by Gen Agent Trust Hub on May 6, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches the @midscene/harmony package from the official npm registry using npx to perform automation tasks. This is a standard method for executing Node.js-based tooling.
  • [COMMAND_EXECUTION]: Executes system commands using the HarmonyOS Device Connector (hdc) to control and interact with hardware devices. These commands are localized to the connected device environment.
  • [DATA_EXFILTRATION]: Captures screenshots of the connected device to perform visual automation. This data is transmitted to the user-configured AI model provider (e.g., Google, Alibaba) for processing, which is the intended core functionality of the vision-driven skill.
  • [PROMPT_INJECTION]: The skill processes visual data from device screens, which constitutes an indirect prompt injection surface. Malicious content displayed on a device screen could theoretically influence the agent's actions, but this is a documented risk of multimodal automation tools.
  • Ingestion points: Device screenshots captured via the take_screenshot command in SKILL.md.
  • Boundary markers: No specific delimiters or boundary markers are used for the screenshot data, which is standard for image-based inputs.
  • Capability inventory: Access to system shell via hdc, file system access for reports, and network access for package downloads.
  • Sanitization: Relies on the security filters and grounding capabilities of the configured AI model provider.
Audit Metadata
Risk Level
SAFE
Analyzed
May 6, 2026, 01:05 PM