native-devtools-mcp-automation

Warn

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill facilitates the download and execution of software from third-party sources not associated with the skill's primary vendor or a recognized trusted organization.
  • Evidence: Installation instructions recommend npx -y native-devtools-mcp and git clone https://github.com/sh3ll3x3c/native-devtools-mcp.
  • [REMOTE_CODE_EXECUTION]: The skill uses npx to fetch and execute a remote package at runtime, which is a high-risk operation when the source is not a trusted vendor.
  • Evidence: npx -y native-devtools-mcp used for both quick start and client configuration.
  • [COMMAND_EXECUTION]: The skill provides the agent with tools to launch arbitrary processes and simulate system-wide keyboard input, which can be leveraged to execute shell commands or malicious scripts.
  • Evidence: launch_app(app_name: string, args?: string[]) and type_text(text: string) tools.
  • [DYNAMIC_EXECUTION]: The skill includes functionality to execute arbitrary JavaScript within a browser or Electron instance, bypassing standard environment restrictions.
  • Evidence: The cdp_eval(expression: string) tool allows for dynamic code execution via the Chrome DevTools Protocol.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a high vulnerability surface for indirect prompt injection due to its computer-vision and DOM-parsing capabilities.
  • Ingestion points: The skill ingests untrusted data through take_screenshot (processed by OCR in find_text), take_ax_snapshot (reading accessibility labels), cdp_take_dom_snapshot (web DOM content), and adb_screenshot (Android UI content).
  • Boundary markers: Absent. There are no instructions or delimiters defined to help the agent distinguish between user-provided system instructions and untrusted text found in the UI or web pages.
  • Capability inventory: The skill possesses extensive capabilities including click, type_text, press_key, launch_app, cdp_eval, cdp_click, and adb_tap.
  • Sanitization: Absent. There is no evidence of sanitization, filtering, or escaping of the content ingested from the screen or browser before it is processed by the agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 8, 2026, 11:44 PM