CVE-2026-72280

CVE-2026-72280 is a KVM arm64 vulnerability where a WARN placed in hypervisor context fires on perfectly valid guest behavior and triggers hyp_panic, crashing the host. The trigger: when FEAT_NV2 (nested virtualization v2) is advertised to an L1 guest, writes to ZCR_EL2 are legitimate. The hypervisor's safety check treated this as unexpected state and invoked the only failure mode available in hyp context — a kernel panic. The guest wasn't doing anything wrong; the hypervisor's instrumentation was miscalibrated. This wasn't a regression introduced by a later change. The WARN was committed alongside FEAT_NV2 support itself, meaning the predicate has been wrong since day one — undetected because KVM's CI pipeline almost certainly exercises L2 guests without FEAT_NV2 exposed to L1, which is the exact configuration that triggers this bug. The test matrix gap isn't accidental; nested virtualization feature interactions are combinatorial and expensive to cover exhaustively. The fix (removing the WARN) is correct, but it creates an uncomfortable side effect: the WARN was a tripwire. Removing it eliminates the signal that would catch other miscalibrated hyp-context checks. The broader risk is that similar false-positive WARNs likely exist elsewhere in nested virtualization paths — predicates written against a mental model of valid guest behavior that FEAT_NV2 (and other ARMv8.x features like SME, MTE, TLBIRANGE) have since expanded. The architectural question is whether WARN-as-panic needs revision. The practical answer for now: audit your KVM kernels for any hyp-context WARNs that depend on guest configuration state, and prioritize testing nested virtualization with FEAT_NV2 advertised to L1. The absence of a regression test for this specific configuration is a structural gap worth closing. If you manage KVM hosts running arm64 with nested virtualization enabled, treat this as a high-priority patch — the exploit condition requires no privilege escalation, just a guest doing what it's permitted to do.

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