NVIDIA NVSentinel: unauthenticated gRPC health-event interface on a world-writable socket allows denial of service
Impact
The platform-connectors component exposes its gRPC health-event interface over a Unix domain socket that is world-writable and performs no authentication. Any local process that can open the socket can submit health events, and a successful exploit leads to denial of service. NVSentinel is the component that decides a GPU node is unhealthy, so a caller that can feed it events is interfering with the fleet's own fault-detection and remediation path - either silencing it or driving it. NVIDIA scores this with a changed scope, reflecting that the damage lands outside the component itself. Confidentiality and integrity are not listed as affected.
Who can reach it
A local process on the node that can reach the socket. NVIDIA's vector requires high existing privileges (AV:L/PR:H), so the realistic case is a compromised privileged daemon or a sidecar running with host filesystem access, not an ordinary tenant pod.
What to do
Apply the NVSentinel update from NVIDIA bulletin 2026/5890 and restart the daemon on each node; the bulletin is the authority on the fixed version. This is a daemon restart rather than a drain - but while NVSentinel is down the fleet is not detecting GPU faults, so roll it through nodes rather than fleet-wide at once. Until then, tightening the socket's permissions and the set of containers given host filesystem access reduces who can reach it.
References
Related entries
- NVIDIA Triton Inference Server: Manipulating the Python backend's shared memory region produces an out-of-bounds readCVE-2025-23333 · NVIDIA Triton Inference ServerMedium
- NVIDIA Triton Inference Server: A crafted request causes an out-of-bounds read in the Python backend, disclosing memoryCVE-2025-23334 · NVIDIA Triton Inference ServerMedium
- HGX / DGX B300: Authentication bypass in the hardware management interfaceCVE-2025-33242 · HGX / DGX B300Medium
- NemoClaw: SSRFCVE-2026-24231 · NemoClawMedium
- Triton Inference Server: Use-after-free in request processingCVE-2026-24266 · Triton Inference ServerMedium
- vGPU Manager: Host impact via GPU command-buffer overflowCVE-2026-24201 · vGPU ManagerMedium
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.