A high-severity security vulnerability discovered within KVM, the native hypervisor built directly into the Linux kernel, recently allowed security researchers to compromise a shared cloud environment and read approximately 127 gigabytes of physical host memory from inside an isolated virtual machine. The flaw, which has been officially tracked and cataloged under the identifier CVE-2026-99312, specifically impacts cloud hosts running on AMD processors that have nested virtualization enabled. While upstream security patches have already been developed and deployed by major infrastructure providers, the discovery underscores the lingering risks associated with complex hardware virtualization extensions and the unintended consequences of legacy security patches.

The vulnerability was initially unearthed by Oren Yomtov, a security researcher operating at the firm Terra Security. Following standard responsible disclosure practices, Yomtov reported the critical flaw through DigitalOcean’s bug bounty program. Upon receiving the notification, the cloud provider quickly verified the exploit on its internal infrastructure, confirming that a user with administrative privileges inside a virtual machine could successfully bypass standard hypervisor isolation boundaries. Recognizing the severity of the issue, the provider treated the matter as critical and worked swiftly alongside Linux kernel maintainers.

Technical communications regarding the vulnerability began in earnest when the researchers notified Linux KVM maintainers on October 21. Following rapid collaboration, a definitive code fix was officially merged into the upstream kernel on November 2. Just a day before the vulnerability details were made public, fixed kernels were pushed upstream. Major cloud providers similarly moved fast to protect their environments; DigitalOcean patched its entire fleet using live-patching mechanisms, assuring its Droplet customers that no manual intervention or corrective action was required on their part. However, independent system administrators and operators managing their own self-hosted KVM environments

on AMD hardware face a different reality. They must either update their underlying host kernels or explicitly disable nested virtualization to mitigate the risk of exploitation.

Linux KVM Flaw CVE-2026-99312 Lets Guest VMs Access Host Memory on AMD Hosts

To demonstrate the gravity of the flaw, the research team conducted controlled testing environments. In one test executed on a DigitalOcean Droplet, the proof-of-concept exploit immediately returned the host system’s kernel version banner. This was followed by the extraction of approximately 127 gigabytes of the physical host’s RAM, mapped entirely as read-only. In a separate test conducted on the researchers’ own bare-metal AMD server—completely isolated from other cloud tenants—the guest virtual machine successfully wrote 32 carefully selected bytes directly into a live page of the host kernel’s memory. A specially loaded witness module running inside the host kernel subsequently confirmed that the unauthorized memory modification had indeed taken place.

Despite the profound implications of a successful guest-to-host memory read and write capability, executing the attack requires a specific starting position. An attacker cannot exploit CVE-2026-99312 merely via network connectivity or external access to a target virtual machine. Instead, an adversary must already possess root-level access from within a compromised or controlled guest virtual machine residing on a vulnerable host.

Crucially, not all cloud environments are at risk. KVM hosts powered by Intel processors remain entirely unaffected by CVE-2026-99312. This is because the vulnerable code path exists exclusively within KVM’s implementation of nested virtualization support for AMD hardware architectures. Furthermore, many enterprise cloud service providers surveyed by the researchers were found to be safe simply because they had chosen not to enable nested virtualization capabilities on their AMD infrastructure by default.

Industry reaction and tracking platforms indicate that the exploit has remained confined to controlled research environments. VirtFusion, a prominent virtualization platform vendor closely monitoring the issue for its base of operators, reported no signs of active exploitation in the wild. Furthermore, no functional exploit code has been published publicly, and security teams currently lack reliable forensic heuristics or detection guidance to identify past attempts at exploitation.

Discrepancies have emerged regarding the formal severity scoring of the vulnerability across different security organizations. The Linux kernel’s internal CVE team assigned the flaw an 8.8 out of 10 CVSS rating, emphasizing a high impact across confidentiality, integrity, and availability. Conversely, Red Hat’s preliminary security assessment adopted a more conservative stance, rating CVE-2026-99312 as Moderate with a score of 5.6, categorizing its potential impact strictly as a denial-of-service vector. Meanwhile, cloud infrastructure providers hosting shared multi-tenant environments viewed the flaw through a far more critical lens due to the immediate threat of cross-tenant data exposure.

Linux KVM Flaw CVE-2026-99312 Lets Guest VMs Access Host Memory on AMD Hosts

A comprehensive technical analysis published by Terra Security breaks down the exact mechanics behind the guest-to-host escape vector. Under the hood, KVM strictly isolates each virtual machine using a nested page table, a specialized translation structure that maps a virtual machine’s memory addresses directly into the host system’s real physical memory. The processor locates this critical table via a pointer maintained independently for every virtual CPU.

Certain advanced configurations allow a virtual machine to host and run a nested virtual machine of its own. When a hardware reset signal known as INIT halts the nested virtual machine, the KVM hypervisor fails to correctly restore the pointer belonging to the outer virtual machine. Consequently, this omission leaves the pointer dangling, referencing a page of memory that the host kernel has already unmapped and freed for other uses.

Once the host kernel inevitably reuses that freed memory page, a malicious or compromised guest can populate the location with arbitrary entries of its own design. The host processor subsequently interprets these forged entries as legitimate page tables. Because the guest can name any physical memory page on the host through these manipulated entries, it gains unauthorized read and write access to vast tracts of server RAM that were never allocated to it. The upstream code fix directly addresses this architectural oversight by forcing KVM to properly reset and reinitialize the memory-management state upon every INIT signal.

Intriguingly, the vulnerability itself was an unintended byproduct of an earlier security fix. According to the research team, a patch deployed in 2022 to resolve a completely separate nested-virtualization bug tracked as CVE-2022-3344 inadvertently introduced the flaw. While that historical patch successfully switched a reset path back to the outer virtual machine’s control block, it omitted the vital page-table pointer reset. As a result, the latent programming error persisted undisturbed within the core Linux kernel for just under four years. This discovery echoes previous findings by the same researchers in their broader "PatchSlop" initiative, which previously demonstrated how remediation efforts can occasionally introduce new kernel vulnerabilities.

Mitigating CVE-2026-99312 requires direct action at the host kernel level, as updating the internal operating systems running inside individual virtual machines offers no protection whatsoever. A host will continue to execute vulnerable code until its underlying kernel is updated and the system is rebooted.

Linux KVM Flaw CVE-2026-99312 Lets Guest VMs Access Host Memory on AMD Hosts

Upstream maintainers officially incorporated the fix into stable and long-term release lines, including stable kernel 7.3.5, along with long-term kernels 6.18.64 and 6.12.118. Because the vulnerable code dates back several years—having been introduced into the 6.1 kernel in 2022 and backported to older releases like 5.15.81 and 6.0.11—a wide array of legacy and current distribution kernels are potentially exposed.

Enterprise Linux distributions and community-driven operating systems have begun rolling out advisories, though many vendor-supplied packages are still in the process of building and deploying corrected kernels. Until a patched host kernel is fully installed and the underlying hypervisor hardware is rebooted, system administrators have few options beyond disabling nested virtualization entirely. While turning off nested virtualization completely closes the vulnerability vector, it simultaneously disables the ability for tenants to run nested virtual machines, a trade-off that cloud operators must communicate clearly to their affected users.

Because nested virtualization is often enabled by default within the Linux kernel’s kvm_amd module, unpatched hosts remain exposed unless administrators have manually altered the configuration. CVE-2026-99312 represents only the latest in a string of prominent KVM isolation failures brought to light this year, following the disclosure of a 16-year-old flaw in KVM shadow memory management impacting both Intel and AMD architectures, as well as the critical PwnEscape vulnerability in Linux networking code that exposed host memory within live sandbox environments.