Executive brief
vLLM is a high-performance engine used for serving artificial intelligence models. A security flaw in its remote communication component allows an attacker to execute arbitrary commands on the server hosting the AI model. This could lead to a complete system takeover, theft of proprietary AI models, or unauthorized access to sensitive customer data processed by the engine.
Technical details
A deserialization vulnerability exists in vLLM version 0.6.0 and earlier within the AsyncEngineRPCServer component. The vulnerability is located in the run_server_loop() function, which calls _make_handler_coro(). This function utilizes cloudpickle.loads() to process incoming messages without any prior validation or sanitization. An unauthenticated remote attacker can exploit this by sending a specially crafted malicious pickle payload to the RPC server entrypoint, leading to arbitrary code execution with the privileges of the vLLM process. As of the advisory date, no patched version is specified.
Affected products
- vllm-project vLLM <= 0.6.0
Timeline
- 2025-03-20: disclosed
- 2025-03-20: advisory