AMD Issues Urgent Fixes for Critical "Mior" and "Skyfall" Flaws Compromising Confidential Computing Virtual Machines

Advanced Micro Devices (AMD) has released security advisories and firmware update timelines to address two newly disclosed attack methods that bypass the robust security guarantees of AMD SEV-SNP (Secure Encrypted Virtualization-Secure Nested Paging). The vulnerabilities allow a malicious host or hypervisor to compromise confidential virtual machines (VMs) that are specifically designed to remain entirely isolated from the underlying server operator.

The two distinct attack vectors, identified as Mior and Skyfall, were made public following extensive research and laboratory testing conducted by academic and industry security teams. While both attacks require an adversary to hold high-level control over the server environment, they successfully punch holes through hardware-enforced memory encryption and access controls, exposing sensitive data, cryptographic keys, and internal registers that are supposed to stay completely private.

Understanding the Mior Attack on SEV-SNP Start-Up

The Mior attack was discovered and demonstrated by a security research team from ETH Zurich, who targeted AMD SEV-SNP while the system was in the critical process of booting up. SEV-SNP is a core technology behind modern confidential computing on AMD EPYC server processors, utilizing hardware-enforced memory encryption to protect workloads running inside virtual machines even from malicious entities managing the host system.

During the initial start-up phase of an SEV-SNP protected environment, a dedicated security co-processor embedded directly on the silicon—known as the Platform Security Processor (PSP)—populates a critical tracking structure called the Reverse Map Table. This table acts as a ledger, recording precisely which entity owns each page of system memory and outlining the specific permissions associated with those pages.

In their laboratory testing on an EPYC 9135 processor belonging to the Zen 5 family, the ETH Zurich researchers devised a clever method to subvert this initialization process. They redirected the PSP’s memory write operations toward a controlled peripheral device, specifically a graphics card connected via PCIe, rather than allowing the writes to settle into the intended main memory locations.

The researchers utilized the graphics card because they theorized that the PSP would subsequently read the values back to verify their integrity, and the controlled hardware could faithfully return the expected data. This manipulation allowed the system start-up sequence to complete without triggering errors, but it left a vital portion of the actual Reverse Map Table—including the very entry responsible for protecting the table itself—completely uninitialized.

Exploiting the Uninitialized Memory Protections

Leaving a portion of the Reverse Map Table uninitialized created a profound security lapse. It granted the host hypervisor the ability to rewrite the protections governing the table. Armed with this capability, the ETH Zurich team enabled debugging features on a confidential virtual machine that had originally been configured with debugging explicitly disabled.

Once debugging was enabled, the researchers leveraged the Platform Security Processor’s native debugging commands to freely read and write the memory space of the target virtual machine. According to the research paper, this exploit achieved a success rate of one hundred percent across one thousand consecutive attempts in their test runs.

Furthermore, the Mior attack tackled the sophisticated attestation mechanisms built into SEV-SNP. Typically, the system measures a virtual machine’s software image during boot and generates cryptographically signed attestation reports via the PSP to confirm the integrity of the running software. In secondary tests, the research team successfully copied the measurement profile of a harmless, trusted virtual machine and transplanted it onto a malicious virtual machine. This trick caused the attestation reports to falsely certify that the malicious workload was actually the benign image.

Executing the Mior attack requires deep privilege, demanding control over both the hypervisor and the system’s BIOS, also referred to as UEFI, alongside a specialized PCIe device like a graphics card. For the purposes of their academic testing, the researchers modified the BIOS firmware to ensure the chip’s internal memory-routing configurations remained unlocked. They noted that this aligns directly with AMD’s established threat model for SEV-SNP, which explicitly assumes that the UEFI environment cannot be trusted by the virtual machine.

The Skyfall Attack and Transient Scheduler Vulnerabilities

While Mior targets the boot sequence and memory mapping of SEV-SNP, the second major threat disclosed this week, known as Skyfall, stems from a collaborative effort between researchers from ETH Zurich and Google. Skyfall builds directly upon Transient Scheduler Attacks, a classification of hardware vulnerabilities affecting AMD processors that can inadvertently leak sensitive data through subtle instruction timing variations. The original transient scheduler flaws were first brought to light in July 2025.

The Skyfall team discovered that when a malicious host or hypervisor deliberately forces a virtual machine to fail during its boot process, the underlying processor core has already cached a portion of the virtual machine’s saved architectural state. Crucially, this vital data is left lingering in internal cache structures without undergoing the standard security cleanup procedures. An adversary operating the hypervisor can then leverage transient execution timing behaviors to read that residual data.

Through this methodology, the researchers successfully recovered saved register values as well as a sensitive disk-encryption key from a test virtual machine. To extract the encryption key, the team modified the virtual machine’s kernel to intentionally pause execution at the exact moment the key resided within the processor registers. The researchers pointed out that in a real-world scenario, a hostile hypervisor could employ an existing technique known as SEV-Step to meticulously advance execution step-by-step until the target material reaches accessible registers.

Unlike Mior, which requires hardware tampering or BIOS modification to manipulate memory routing, Skyfall can be executed entirely through software control of a malicious host or hypervisor, without altering physical hardware components. The researchers successfully demonstrated the technique on an EPYC 7313 Zen 3 processor, as well as EPYC 9124 and EPYC 9334 Zen 4 chips, though they did not observe the data leak on Zen 5 architecture.

The Skyfall research team also extended their work beyond virtual machines, demonstrating that privileged software on a host machine could read decrypted microcode updates across all three tested chip families. They were additionally able to recover register values stored within System Management Mode memory, a highly privileged operational state that is normally inaccessible even to the standard operating system kernel.

However, the program committee for the IEEE Symposium on Security and Privacy noted that mounting a practical attack against virtual machines via Skyfall presents significant hurdles. A successful exploitation demands a dedicated processor core free from competing activity, the ability to induce failing virtual machine boots alongside high-speed performance counter reads, and a virtual machine that maintains stability without crashing or timing out during the iterative attack window.

Affected Processors, Vulnerability Tracking, and Manufacturer Responses

AMD has formally cataloged Mior under the tracking identifier CVE-2026-0462, assigning it a CVSS 4.0 severity rating of 5.9 in an official security advisory. The company designated only the SEV-SNP architecture as vulnerable to Mior, leaving earlier iterations like SEV and SEV-ES unaffected. For Skyfall’s virtual machine memory leakage, AMD issued a separate advisory tracking the issue as CVE-2026-40671 with a severity score of 4.9, detailing how the flaw exposes internal virtual machine register data.

The chipmaker outlined remediation plans affecting five major generations of EPYC server processor families, spanning the 7003, 8004, 8005, 9004, and 9005 product lines. AMD released initial SEV firmware updates for Mior and microcode updates for Skyfall on August 12. Additional platform firmware packages are scheduled for wider release, with updates intended to roll out across server manufacturers through November for Skyfall mitigations and December for Mior platform fixes, while embedded processor variants are slated to receive patches between January and April 2027.

Because these essential mitigations are integrated directly into host-level firmware and BIOS configurations, enterprise customers who rent confidential virtual machines rely heavily on their respective cloud service providers and server vendors to apply the patches. Major hardware manufacturers, including Supermicro, have begun publishing specific BIOS update roadmaps covering motherboards across their server product portfolios, while reminding administrators that baseboard management controller firmware updates are also required to fully secure the platforms.

Independent security bulletins have also addressed related risks, such as an EDK II project advisory tracking a System Management Mode vulnerability under CVE-2026-18448 with a 5.6 severity rating, which stems from missing transient scheduler mitigations in firmware source code. As hardware vendors work to distribute the necessary microcode, SEV firmware, and BIOS updates, security researchers emphasize that maintaining rigorous patch management across host hypervisors and system firmware remains vital to preserving the integrity of confidential computing environments.