New Security Flaws Target AMD SEV-SNP Confidential Computing and Processor Architecture

Two newly disclosed attack vectors have been brought to light, demonstrating how malicious hosts can read sensitive data inside AMD SEV-SNP-protected virtual machines (VMs), which are fundamentally engineered to keep server operators and cloud administrators out.

The security findings expose vulnerabilities that cloud computing providers and enterprise environments must address through upcoming firmware and microcode updates. Researchers have demonstrated both attacks in lab testing environments utilizing AMD EPYC server processors, prompting the hardware manufacturer to issue security advisories and a comprehensive rollout schedule for patches.

One of the newly revealed techniques, designated as "Mior," possesses the capability not only to read data from a virtual machine but also to modify its memory and forge the signed cryptographic reports that confirm which specific software workload is running. Successfully executing Mior requires high-level control over the server’s hypervisor as well as direct command of the server’s BIOS, also known as UEFI.

The second attack vector, known as "Skyfall," builds upon previously identified transient execution flaws and requires a malicious host or hypervisor without necessitating physical hardware tampering.

Understanding AMD SEV-SNP and the Mior Attack Vector

Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) is a core technology designed to bring advanced confidential computing capabilities to AMD EPYC server processors. By utilizing hardware-enforced memory encryption and strict access control mechanisms, SEV-SNP is built to protect virtual machines from unauthorized inspection or tampering, even by entities possessing hypervisor-level privileges on the host machine.

The academic research team behind the Mior attack—originating from ETH Zurich—targeted SEV-SNP specifically during its initial boot-up phase. During this critical start-up sequence, a dedicated hardware security co-processor embedded directly on the silicon chip, known as the Platform Security Processor (PSP), populates an internal structure called the Reverse Map Table, or RMP. The RMP functions as an essential ledger that records the ownership of every individual page of memory and dictates what operations are legally permitted to be performed on them.

During their laboratory experiments, the research team successfully redirected the PSP’s write operations away from main memory and into the memory space of a controlled graphics card. The researchers utilized a discrete PCIe device because they theorized that the PSP would subsequently read the values back to verify their integrity, and the external card was capable of returning the expected data. Consequently, the system initialization completed without triggering apparent errors, but a crucial portion of the genuine RMP—including the specific entry tasked with protecting the table itself—remained uninitialized.

This anomalous state allowed the hypervisor to rewrite the protective boundaries of the RMP. Following this manipulation, the team enabled debugging capabilities on a confidential virtual machine that originally had debugging disabled. They then leveraged the PSP’s native debugging commands to freely read and write to the VM’s protected memory space. According to the researchers, this sophisticated exploitation method succeeded consistently across one thousand consecutive test runs.

New Mior and Skyfall Attacks Let Rogue Cloud Hosts Steal Secrets From AMD SEV-SNP VMs

Furthermore, SEV-SNP measures a virtual machine’s software image during start-up, and the PSP signs attestation reports that incorporate this cryptographic measurement. In a secondary phase of testing, the research team copied the software measurement of a benign, harmless virtual machine into a malicious virtual machine. This trickery forced the malicious VM to generate attestation reports that misleadingly described the harmless software image instead of its own contents.

The Mior attack requires administrative control over both the hypervisor and the UEFI/BIOS, alongside a peripheral PCIe device such as a graphics card. For the purpose of their laboratory validation, the researchers modified the BIOS configuration to leave the processor’s internal memory-routing settings unlocked. Emphasizing the parameters of their threat model, the team noted that they operated strictly within AMD’s defined parameters for SEV-SNP, where the host hardware explicitly distrusts the UEFI, thereby acknowledging that an attacker might modify it.

The research team executed the Mior attack successfully on an EPYC 9135 processor, which is based on the Zen 5 architecture. They noted that older Zen 3 and Zen 4 chips are also potentially affected, though they were unable to test them directly due to hardware limitations preventing BIOS modifications on those specific platforms. Notably, the debugging and attestation exploits remained fully functional even when optional ciphertext hiding was enabled—a security feature designed precisely to prevent hypervisors from reading encrypted VM memory contents.

The researchers pointed out that Mior is entirely distinct from earlier RMP-corrupting vulnerabilities uncovered by overlapping teams, such as the Fabricked attack and the RMPocalypse vulnerability, and it successfully functioned on older firmware versions predating those previous patches.

The Skyfall Attack and Transient Execution Vulnerabilities

The second major disclosure, Skyfall, was developed by a collaborative team of researchers from ETH Zurich and Google. This attack leverages Transient Scheduler Attacks (TSA), a broader class of architectural vulnerabilities in AMD processors capable of leaking confidential information through precise instruction timing. TSA-related flaws were initially disclosed to the public in July 2025.

The investigative team discovered that when a hypervisor intentionally triggers a scenario where a virtual machine fails to boot up correctly, the processor has already cached a portion of that VM’s saved execution state. Crucially, this sensitive data is left lingering in the cache without undergoing standard security clearing procedures, allowing a malicious hypervisor to extract it via TSA techniques.

Through this methodology, the researchers successfully recovered saved register values from a virtual machine alongside a critical disk-encryption key. To isolate the encryption key, the team modified the guest VM’s operating system kernel to temporarily pause execution at the exact moment the key resided within the processor registers. In a realistic attack scenario, researchers explained that a hypervisor could achieve similar results by utilizing a technique known as SEV-Step, systematically advancing the virtual machine instruction by instruction until the target key surfaces in the registers.

While Skyfall can read VM memory, its reach is structurally restricted. The attack is limited to recovering data from three specific fixed cache lines located within a targeted memory page—specifically at offsets 0x0c0, 0x380, and 0x3c0—leaving the remainder of the page inaccessible.

The Skyfall attack requires a malicious host or hypervisor but does not involve physical hardware tampering. The team validated the technique across multiple generations of silicon, including EPYC 7313 processors based on Zen 3 architecture, as well as EPYC 9124 and 9334 chips based on Zen 4 architecture. However, the data leakage could not be replicated on Zen 5 processors.

New Mior and Skyfall Attacks Let Rogue Cloud Hosts Steal Secrets From AMD SEV-SNP VMs

Beyond virtual machines, the Skyfall researchers demonstrated the ability to read decrypted microcode updates across all three tested chip generations. They also recovered register values stored within System Management Mode (SMM) memory, a highly privileged operational state normally inaccessible even to operating system kernels. Both of these secondary attack vectors require privileged software execution on the underlying machine.

Reviewing the academic implications, the program committee for the IEEE Symposium on Security and Privacy—which accepted the formal Skyfall research paper—noted that practical exploitation against VMs in real-world deployments would face considerable hurdles. Successful execution requires a processor core free from competing background activity, the ability to induce failing VM starts and read performance counters at extremely high speeds, and a virtual machine that avoids crashing or timing out prematurely during the attack sequence.

Affected Processors, Vulnerability Ratings, and Remediation Efforts

AMD is tracking the Mior vulnerability under the identifier CVE-2026-0462, assigning it a moderate severity rating of 5.9 on the CVSS 4.0 scale within an official security advisory. Meanwhile, Skyfall’s VM data-leakage component is tracked as CVE-2026-40671 with a CVSS rating of 4.9, detailed in a secondary advisory that outlines how the vulnerability exposes virtual machine register information.

AMD’s advisories confirm that Mior impacts SEV-SNP implementations, while excluding the older, legacy SEV and SEV-ES iterations. Both vulnerabilities span the same five major EPYC server processor families, encompassing the 7003, 8004, 8005, 9004, and 9005 series.

To counter these threats, AMD has structured a phased update schedule. SEV firmware updates addressing Mior and microcode updates addressing Skyfall were dated and released on August 12. Corresponding platform and motherboard firmware updates from hardware manufacturers are scheduled for staggered rollouts, with Skyfall-related platform firmware targeting a November release, Mior-related platform firmware slated for December, and patches for embedded chip variants scheduled for deployment between January and April 2027. Additionally, specialized hardware such as the Instinct MI300A accelerator received a dedicated Skyfall patch in September.

Because these security mitigations rely heavily on host-level firmware and BIOS updates, customers who rent confidential virtual machines are heavily dependent on their respective cloud service providers to apply the patches promptly. Furthermore, hardware manufacturers such as Supermicro have begun publishing dedicated BIOS fix lists covering vulnerable motherboards across their enterprise server product lines, while noting that supplementary Baseboard Management Controller (BMC) firmware updates will also be required.

Separately, the EDK II open-source project assigned Skyfall’s SMM memory leakage issue the identifier CVE-2026-18448 with a CVSS score of 5.6. An accompanying AMD security bulletin traces this specific flaw back to missing TSA mitigations within foundational EDK II firmware codebase elements. This bulletin advises administrators of affected EPYC and Ryzen platforms to apply preceding TSA mitigations alongside upcoming EDK II software patches.

Neither AMD nor the participating academic research teams have reported any instances of these vulnerabilities being actively exploited in the wild prior to their public disclosure. Proof-of-concept exploit code has been published by the researchers to encourage defensive auditing, though successful execution of the Mior proof-of-concept requires a specially modified BIOS environment with unlocked memory-routing configurations. Enterprises utilizing AMD EPYC confidential computing infrastructure are strongly advised to coordinate with their hardware vendors and cloud service providers to verify firmware release schedules and apply the necessary security mitigations as they become available.