CVE-2026-72284

CVE-2026-72284 is a KVM host kernel denial of service that triggers when a guest disables paravirtualized EOIs at a specific point in the vCPU run loop. The bug manifests as a kernel panic in kvm_lapic_sync_from_vapic during state synchronization between guest execution and hypervisor management—an assertion fires because KVM finds itself in a logically impossible configuration: a PV EOI request is pending but the feature has already been disabled by the guest. The vulnerability is a guest-to-host DoS that takes down every VM on the affected host, not just the malicious one. The CVSS 7.1 is appropriate for the mechanism but undersells the multi-tenant blast radius—in a shared hypervisor environment, one guest's crash creates collateral damage across unrelated workloads. The patch adds a guard that checks whether PV EOIs are still enabled before acting on a pending request. This prevents the crash but doesn't fix the underlying architectural gap: KVM's internal request mechanism can outlive the conditions that made it valid. The request was designed as a performance optimization—to eliminate VM exits—but that optimization creates state that exists only until the next exit checks it. When the guest changes the feature flag during that window, the hypervisor's internal state becomes stale. The guard pattern is the key indicator for defenders: this is the same fix approach that has historically been applied to similar TOCTOU patterns in KVM's paravirtualization interfaces. Each guard silences a specific crash without restructuring how KVM manages request validity relative to guest-controlled state transitions. The implication is that other PV interfaces likely contain the same latent inconsistency—every paravirtualization path that touches guest-controlled flags and uses async requests is operating on the same undocumented assumption that those flags won't change mid-cycle. For assessment: prioritize patching, but also audit other paravirtualization interfaces for similar TOCTOU windows. The exploitability is low—the race window is bounded by VM exit/run cycles and requires a privileged guest—but the impact is high in multi-tenant environments. Assume the architectural gap persists, assume similar patterns exist elsewhere, and measure your real exposure not as 'can this guest crash the host' but as 'when this host crashes, how many unrelated workloads go down with it.'

Reviewed through automated stages and approved by a human before publication.