GPU VulnDB

Database/AI/ML frameworks & serving

NVIDIA NemoClaw: OS command injection in the NIM management component

CVE-2026-65090AI/ML frameworks & servingcurated

Impact

NemoClaw's NIM management component passes attacker-influenced input into an OS command, giving code execution, data tampering, information disclosure and denial of service. This is the component that starts, stops and configures NIM inference microservices, so it typically runs with enough privilege to manage containers and GPU devices on the node - command execution there inherits that reach. An attacker who lands it can tamper with which model artifacts get served, read whatever credentials the management layer holds for pulling images, or shut the serving stack down. NVIDIA rates it 7.8 with high impact across confidentiality, integrity and availability.

Who can reach it

A local, low-privileged actor able to feed input to the NIM management component (CVSS AV:L, PR:L). Authentication is required; no user interaction.

What to do

Update NemoClaw from the NVIDIA/NemoClaw GitHub repo. Bulletin 5872 lists versions 0 through 0.0.25 as affected with fixed builds varying per CVE - take the exact version from the bulletin's Security Updates table rather than assuming the latest tag covers this one. Rolling the update means restarting the NemoClaw services and the NIMs they manage, so the node's inference capacity drops for the restart window.

References

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.