Unsafe Deserialization in NVIDIA TensorRT-LLM MPI Server

Successful exploitation may allow an attacker to achieve code execution in the context of the vulnerable MPI server component. The provided content also states that exploitation can cause denial of service, data tampering, and information disclosure, indicating potential compromise of confidentiality, integrity, and availability.

Where immediate patching is not possible, reduce exposure of the vulnerable MPI server by limiting network reachability and restricting access to trusted hosts and users only. In multi-GPU deployments, the provided content specifically recommends using the trtllm-llmapi-launch executable directly to reduce network exposure. Additional compensating controls include isolating TensorRT-LLM services, minimizing untrusted input to the MPI server, and monitoring for abnormal activity targeting message-handling interfaces.

Upgrade NVIDIA TensorRT-LLM to version 1.2.1 or later, as recommended in the provided content. Apply NVIDIA’s vendor-supplied fixes across affected deployments and ensure the MPI server component is updated wherever TensorRT-LLM is installed.