Skip to content
Xianzhi Community (Aliyun)·Indexed by this site · Originally published· Event →· Its place in the scan record →

CVE-2026-30623 vulnerability analysis: reproduction and source code audit of the LiteLLM Proxy MCP interface command injection vulnerability

AI summary

The author audited and reproduced the command injection vulnerability CVE-2026-30623 in the LiteLLM Proxy MCP interface; an attacker holding a valid API Key can execute arbitrary system commands through the /mcp-rest/test/connection interface. When this interface receives a stdio-type MCP Server configuration, it does not validate command or args and passes them directly to a subprocess for execution; an attacker sets command to bash and passes a script via -c, bypassing the injection protection of list-based argument passing. Exploitation requires a valid Bearer API Key; unauthorized requests return 401, the test-connection branch does not use the database, and the vulnerability can still be triggered when the database is offline. The author reasons that an attacker could steal model keys, plant backdoors and move laterally. Remediation suggestions include command allowlisting, fine-grained API Key permissions and subprocess sandbox isolation; before a patch, the MCP feature can be disabled.

Read originalxz.aliyun.com