NVIDIA Container Toolkit CVE-2026-24260

Patch available per vendor advisory - no exact fixed version was included in the input data, so obtain and apply the patched NVIDIA Container Toolkit and GPU Operator release specified in NVIDIA product-security advisory 5850 (https://github.com/NVIDIA/product-security/tree/main/2026/5850) and confirm the fixed version against the NVD entry (https://nvd.nist.gov/vuln/detail/CVE-2026-24260) before deploying; on Kubernetes, upgrade via the GPU Operator so all nodes receive the fixed toolkit consistently. As compensating controls until patched: restrict who can launch GPU containers by tightening RBAC and admission control so only trusted workloads get the nvidia runtime (trade-off: blocks some legitimate GPU jobs), avoid running GPU containers as privileged or with host mounts that overlap toolkit-managed paths (trade-off: may break workloads relying on host device access), and enforce read-only container root filesystems and drop CAP_SYS_ADMIN where feasible to reduce an attacker's ability to stage the symlink/path swap needed to win the race (trade-off: some GPU workloads require specific capabilities). Monitor node logs for unexpected mount/symlink activity around container startup as a detection aid.