NVIDIA Data Center GPU Manager Remote Memory Corruption
Posted on 03 June 2022
NVIDIA DCGM runs on machines with NVIDIA GPUs to gather telemetry and GPU health data. nv-hostengine is a daemon that by default listens on the loopback interface, but can also listen on the network for requests coming in on port 5555 (remote mgmt). A native client named DCGMI allows users to make requests to the daemon to support a variety of functions. Malformed packets can cause the daemon (running as root or user account) to crash or potentially result in code execution. Versions less than 2.3.5 are affected.
