NVIDIA has released six security advisories on 13 October 2026, addressing 18 vulnerabilities across its AI software stack. The most critical entry is CVE-2026-65162, rated CVSS 10.0, which affects NVIDIA AIStore and could allow remote code execution. Eight of the flaws pertain to AIStore, all of which carry a “no privileges required” condition. Overall, there are 2 medium and 14 high-severity vulnerabilities in the suite, with ongoing assessments noting that exploitation in the wild has not been confirmed.
AIStore is NVIDIA’s storage solution for large AI datasets, frequently deployed in Kubernetes clusters hosting GPUs, models and training data; breaches here could expose sensitive intellectual property. The report details how the CVSS 10.0 flaw stems from a legacy ETL API that could let a remote attacker supply dangerous Kubernetes host and security settings, enabling code execution and privilege escalation.
Additional high-severity issues include an unauthenticated token-revocation endpoint that could be used to tamper with data or trigger a denial of service, plus several path traversal, data leakage and PSR-type flaws affecting AIStore’s ETL workflows. Other products in the AI stack—Megatron-LM, Isaac Lab, DALI, NVSentinel and Warp—also contain bugs that could enable remote code execution or tampering if compromised files, inputs or model checkpoints are involved, though none are reported as exploited to date.
NVIDIA’s patch guidance calls for upgrading AIStore to 5.0 and updating Megatron-LM, DALI, Isaac Lab, NVSentinel and Warp to specific newer versions. Practically, operators should also keep AIStore proxies and targets off public networks, restrict API access, and only load model checkpoints from trusted sources. The advisory notes no required credentials for several AIStore flaws, underscoring the importance of timely patching in AI cluster environments.