CVE-2026-99312 is a use-after-free in KVM, the Linux kernel hypervisor, affecting AMD systems with nested virtualization enabled. The flaw arises when a virtual CPU receives an INIT reset while running a nested VM: KVM can skip the code that refreshes the outer VM’s nested-page-table pointer. If the attacker has root inside the guest, they can arrange for KVM to free the page named by that stale pointer. The processor continues using it as the top level of the guest’s memory translation, even after KVM has repurposed it.
The researchers describe turning that stale pointer into a path toward host memory. They caused KVM to reuse the freed page as a lower-level page table, then used guest-controlled memory and the processor’s page-table Accessed bits to identify host addresses belonging to their VM. On a DigitalOcean droplet, they read the host kernel’s version banner and about 127 GB of host RAM, potentially exposing other tenants’ data. On their own bare-metal AMD server, they demonstrated writing 32 chosen bytes into a live host-kernel page; such an arbitrary write is a foundation for code execution, though the described demonstration does not claim to have achieved it. The key reuse step succeeded on six of twelve boots, making the attack unreliable on any single attempt but practical to retry. The bug had been present for nearly four years and was introduced by a security fix.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.