GHSA-54j7-grvr-9xwg · CVE-2025-59834 · 갱신 2026. 9. 10.
| 생태계 | 패키지 | 수정 버전 |
|---|---|---|
| npm | adb-mcp | 없음 |
The argument to the tool, AdbDevicesSchema, is a Zod inferred type defined in the src/types.ts file in the project:
export const inspectUiInputSchema = {
device: z.string().optional().describe("Specific device ID (optional)"),
outputPath: z.string().optional().describe("Custom output path on device (default: /sdcard/window_dump.xml)"),
asBase64: z.boolean().optional().default(false).describe("Return XML content as base64 (default: false)")
};
and exposes device as a string which is an open way to trick the LLM into pushing arbitrary strings into it and hence achieve the command injection exploitation.
When LLMs are tricked through prompt injection (and other techniques and attack vectors) to call the tool with input that uses special shell characters such as ; rm -rf /tmp;# (be careful actually executing this payload) and other payload variations, the full command-line text will be interepted by the shell and result in other commands except of ps executing on the host running the MCP Server.
Reference example from prior security research on this topic, demonstrating how a similarly vulnerable MCP Server connected to Cursor is abused with prompt injection to bypass the developer's intended command:
User initiated and remote command injection on a running MCP Server.
exec. Use execFile instead, which pins the command and provides the arguments as array elements.-- notation to terminate command and command-line flag, and indicate that the text after the -- double dash notation is benign value.Disclosed by Liran Tal