GPU VulnDB

Database/AI/ML frameworks & serving

NVIDIA NemoClaw: inference service comes up without authentication, reachable from the adjacent network

CVE-2026-65105AI/ML frameworks & servingcurated

Impact

NemoClaw's inference server setup leaves the inference endpoint exposed with no authentication required. Anyone who can reach the service on the local network can query it and pull back information, or drive it into denial of service - NVIDIA rates confidentiality and availability high, integrity none. On a GPU node this means prompts, model responses and whatever the endpoint reveals about the loaded model are readable by an unauthenticated neighbour, and the serving process can be knocked over, taking the GPUs it holds out of production until it is restarted. In a shared cluster where pod networks are flat, 'adjacent' includes every other tenant on the same segment.

Who can reach it

Any unauthenticated party with adjacent-network reach to the NemoClaw inference endpoint (CVSS AV:A, PR:N, UI:N). No credentials and no user interaction needed.

What to do

Update NemoClaw from the NVIDIA/NemoClaw GitHub repo; bulletin 5872 lists versions 0 through 0.0.25 as affected, with the fixed version varying per CVE within the 0.0.1-0.0.25 range - check the Security Updates table in the bulletin for the exact build. Applying it means restarting the NemoClaw inference service, so the GPUs it holds drop out for the duration. Until then, restrict network reach to the inference port with a NetworkPolicy or host firewall rather than relying on the service to authenticate callers.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-65105https://github.com/NVIDIA/product-security/tree/main/2026/5872https://www.cve.org/CVERecord?id=CVE-2026-65105

Related entries

All AI/ML frameworks & serving entries

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.