Linux Kernel CVE-2026-68329

Exploitation requires all of the following specific conditions: (1) an AMD processor with AMD-Vi (IOMMU) hardware present and the iommu/amd Linux kernel driver active - Intel IOMMU (VT-d) systems are unaffected; (2) symmetric multi-processing (SMP) with at least two CPUs concurrently issuing IOMMU commands to the same IOMMU instance, as the race is strictly between two CPUs sharing one IOMMU's command queue and need_sync flag; (3) local code execution with at minimum low-privilege (PR:L per CVSS vector) sufficient to trigger concurrent kernel-mode IOMMU invalidation paths; (4) active IOMMU usage, typically via DMA-capable device I/O or device passthrough, to produce the necessary command queue activity. Exploitation is limited by the narrow CPU timing window required to win the race between CWAIT queuing and need_sync testing, the AMD-specific hardware dependency, and the absence of any known automation or public tooling. Remote exploitation is not possible - the CVSS AV:L vector confirms the local-only attack surface.

The vendor-assigned CVSS 3.1 score of 8.8 (AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H) captures the high potential impact of a kernel use-after-free that crosses isolation scope (S:C), but AC:L is arguably optimistic - winning a multi-CPU race against IOMMU hardware completion is a non-trivial timing challenge that better fits AC:H. EPSS at 0.20% (10th percentile) is consistent with unexploited kernel internals bugs requiring specific hardware and timing conditions. The vulnerability is not listed in CISA KEV, and no POC is publicly available. Practical risk is highest in AMD-based cloud and virtualization infrastructure where: (a) IOMMU is actively enforcing DMA isolation between VMs, (b) multiple CPUs are concurrently performing I/O-intensive workloads, and (c) untrusted local code execution is possible. For typical enterprise Linux servers without active AMD IOMMU-backed device passthrough, the exploitation window is narrow. EPSS and absence of KEV/POC together suggest this remains a theoretical priority rather than an actively weaponized flaw, though the kernel stable team's multi-branch backport to six version lines signals they assessed it as a meaningful security boundary violation.

A local attacker with low-privilege shell access on a multi-CPU AMD server where AMD IOMMU is enabled orchestrates two concurrent threads performing IOMMU-backed DMA operations: one thread queues an IOMMU invalidation command and then frees the associated page-table pages, while a second thread races to call iommu_completion_wait() - reading need_sync == false without holding the lock, bypassing the hardware sync, and returning before the CWAIT hardware signal fires. The IOMMU proceeds to walk the now-freed page-table pages, producing a use-after-free condition that can yield arbitrary kernel memory reads or code execution. No public proof-of-concept exists at time of analysis.