Linux Kernel CVE-2026-72288

Requires the host to be an arm64 system running KVM with vGIC/GICv3 ITS LPI support, and the attacker must control a guest VM able to drive interrupt operations on multiple vCPUs concurrently. The concrete trigger is a race between an LPI-disable write to one vCPU's redistributor (RD) and an interrupt-affinity migration of that same LPI to another vCPU, hitting the interval where vgic_prune_ap_list() releases and reacquires the irq and ap_list locks. Limiting factors: exploitation is not possible on x86 or non-KVM hosts, requires an already-present untrusted guest (not remote/unauthenticated despite the input's PR:N), and depends on winning a narrow timing window, making success probabilistic rather than deterministic.

The supplied CVSS 3.1 vector (AV:L/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H = 9.3) overstates ease of exploitation for this class of bug: real exploitation requires winning a three-way race across vCPUs, which in practice is AC:H, not AC:L, and requires the attacker to already control a guest VM (effectively PR:L, not PR:N), so the true score is lower than 9.3. All signals point to low near-term risk: EPSS is 0.21% (11th percentile), there is no KEV listing, and no public exploit identified at time of analysis. The scope-changed rating (S:C) is justified because a guest can corrupt host-kernel memory. The realistic impact is host denial-of-service (UAF panic) with a harder-to-achieve path to host-kernel code execution; the input's 'Information Disclosure' tag is not well supported by the description and should be treated as an approximate label. Net: patch during normal kernel maintenance cycles rather than as an emergency, but multi-tenant arm64 virtualization hosts should prioritize it because the attacker is an untrusted guest and the target is the host kernel - even a mid-sized cloud operator running untrusted arm64 tenants should not defer it indefinitely. Flagged uncertainty: the input CVSS shows PR:N/AC:L, which conflicts with the description's race-and-guest-control requirements - verify privilege and complexity assumptions against the actual attack model before ranking severity.

An attacker who already controls an untrusted guest VM on an arm64 KVM host uses multiple vCPUs to simultaneously disable LPIs on one vCPU's redistributor while migrating the same LPI's affinity to another vCPU, deliberately hitting the lock-drop window in vgic_prune_ap_list(). Winning the race frees the interrupt object while it is still referenced, corrupting host-kernel memory to crash the hypervisor or, with careful heap grooming, attempt host-kernel code execution. No public exploit identified at time of analysis, and the tight timing window makes reliable exploitation difficult.