NVIDIA has released patches for two high-severity vulnerabilities in its Triton Inference Server, addressing security flaws that could allow unauthenticated remote attackers to cause a denial of service and access unauthorised data. The issues, tracked as CVE-2026-16497 and CVE-2026-47625, affect Linux installations across multiple release trains. There is no indication of active exploitation or publicly available PoCs at this time.
CVE-2026-16497 arises from an excessive iteration condition in the Triton server, enabling a remote actor to send malformed inputs that exhaust server resources. CVE-2026-47625 involves missing authorization checks, with the potential to enable information disclosure, data tampering, or denial of service over the network without requiring user authentication. Both flaws are rated with a CVSSv3 base score of 7.5, and are not reported as exploited in the wild to date.
Affected versions include Linux releases from 0.0 up to 26.06 for CVE-2026-16497, and 0.0 up to 26.03 for CVE-2026-47625. NVIDIA advises updating to fixed releases to mitigate risk, with recommended targets of release 26.07 or 26.04. The fixes are published in the Triton Inference Server GitHub repository. Organisations running Triton on Linux should apply the updates promptly to secure production ML deployments across cloud, edge, and data centre environments.