GPU VulnDB

Database/AI/ML frameworks & serving

NVIDIA NemoClaw: OS command injection through the command-line interface

CVE-2026-65099AI/ML frameworks & servingcurated

Impact

The NemoClaw CLI builds an OS command from input it does not adequately sanitise, so a crafted argument or value executes as a shell command with the privileges of whoever runs the CLI. NVIDIA lists code execution, data tampering, information disclosure and denial of service. The practical risk on a GPU node is anything that invokes the CLI on behalf of someone else - a deployment script, a CI job, or an operator running it against tenant-supplied names or paths - because the injected command then runs at the caller's privilege rather than the submitter's. The record does not say the CLI is directly reachable by untrusted users; treat the injection point as whatever supplies its arguments.

Who can reach it

A local, low-privileged actor who can influence the arguments or input passed to the NemoClaw CLI (CVSS AV:L, PR:L). Authentication required.

What to do

Update NemoClaw from the NVIDIA/NemoClaw GitHub repo; versions 0 through 0.0.25 are affected per bulletin 5872, with the fixed build listed per CVE in that bulletin's Security Updates table. Replacing the CLI binary is enough for new invocations - nothing needs a reboot - but any long-running service or automation that shells out to it should be restarted so it picks up the new version.

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.