Two newly disclosed cyberattacks, named Mior and Skyfall, have demonstrated how malicious cloud hosts could bypass protections in AMD’s Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) technology to access sensitive information stored inside protected virtual machines.
The attacks undermine a security feature designed to prevent cloud providers and server operators from accessing the data of customers running confidential virtual machines.
Researchers demonstrated that both techniques could extract information from SEV-SNP-protected virtual machines running on AMD EPYC server processors. While both attacks require control of the host hypervisor, Mior also requires control of the system BIOS or UEFI firmware.
Mior goes further by allowing attackers to modify virtual machine memory and manipulate cryptographically signed attestation reports, which are designed to verify the software running inside a protected virtual machine.
AMD has acknowledged both vulnerabilities and released firmware-related security updates. The company lists SEV firmware dated August 12, 2026, for Mior and microcode updates for Skyfall. Additional platform firmware updates are expected through server manufacturers, with fixes for some embedded processors scheduled for 2027.
Customers are advised to obtain the relevant BIOS updates from their server manufacturers.
AMD SEV-SNP is a confidential computing technology built into EPYC processors. It uses hardware-based memory encryption and access controls to isolate virtual machines from the underlying hypervisor, preventing even the server operator from directly inspecting their protected memory.
However, the newly disclosed research shows that weaknesses in processor initialization and instruction execution can undermine these protections under specific conditions.
How the Mior Attack Works
Researchers from ETH Zurich discovered that Mior exploits a weakness in the way AMD’s security architecture initializes memory protections when a confidential virtual machine starts.
During startup, a dedicated security processor known as the Platform Security Processor (PSP) initializes a data structure called the Reverse Map Table (RMP). The table records which virtual machine owns individual memory pages and determines how those pages can be accessed.
The researchers found a way to redirect the PSP’s memory writes away from the system’s main memory and into memory belonging to a graphics card under their control.
They selected a graphics card because they believed the PSP would read back and verify the values written during initialization, allowing the device to return the expected information.
Although the startup process completed without reporting an error, some entries in the actual RMP remained uninitialized. These included the entry responsible for protecting the table itself.
This allowed the malicious hypervisor to modify memory protection settings that should have been enforced by the hardware.
The researchers subsequently enabled debugging on a confidential virtual machine that had debugging disabled and used commands from the PSP to read and modify its memory.
They reported that the technique succeeded in all 1,000 attempts conducted across two test runs.
Mior also demonstrated a way to manipulate SEV-SNP attestation reports. These reports contain cryptographically signed measurements of the software image loaded when a virtual machine starts.
By copying the measurement of a legitimate virtual machine into a malicious one, the researchers made the malicious environment produce attestation reports that appeared to describe the harmless software image.
According to the Mior research paper, the attack requires control of the hypervisor and BIOS or UEFI firmware, as well as access to a PCIe device such as a graphics card.
The researchers modified the BIOS during their experiments to prevent the processor’s memory-routing configuration from being locked.
They said their research operates within AMD’s stated SEV-SNP threat model, which does not treat UEFI firmware as trustworthy and therefore allows for the possibility of an attacker modifying it.
The team tested Mior on an AMD EPYC 9135 processor based on the Zen 5 architecture. Although the researchers believe Zen 3 and Zen 4 processors may also be affected, they did not test those platforms because they lacked systems that allowed the required BIOS modifications.
The attack also succeeded with SEV-SNP’s optional ciphertext hiding feature enabled. This feature is intended to prevent the hypervisor from reading encrypted virtual machine memory.
The researchers noted that Mior is separate from two previously disclosed attacks, Fabricked and RMPocalypse, which also targeted the RMP during SEV-SNP initialization. They said Mior exploits an independent weakness and remains effective against older firmware versions released before the earlier fixes.
Skyfall Exploits Processor Timing Behaviour
Skyfall was developed by researchers from ETH Zurich and Google. Unlike Mior, it builds on a previously identified category of AMD processor vulnerabilities known as Transient Scheduler Attacks (TSA), which were first disclosed in July 2025.
The researchers discovered that when a malicious hypervisor deliberately forces a virtual machine to fail during startup, portions of its saved processor state may already have been placed in the processor’s cache.
Instead of being properly cleared, the information can remain accessible through timing-based techniques associated with TSA.
Using this weakness, the researchers recovered saved register values and a disk-encryption key from a test virtual machine.
To obtain the encryption key, they modified the virtual machine’s kernel so it would pause at the precise moment the key was stored in processor registers.
They explained that an attacker could potentially achieve the same result in a real environment by using a technique called SEV-Step to control the virtual machine’s execution until the encryption key appears in the registers.
Skyfall can also expose limited portions of virtual machine memory. However, the researchers said the attack is restricted to three specific cache-line positions within a targeted memory page, located at offsets 0x0c0, 0x380 and 0x3c0.
The remaining portions of the page cannot be accessed through this technique.
Unlike Mior, Skyfall does not require modifications to the system’s BIOS or physical hardware. However, it still requires a malicious host or hypervisor.
The researchers tested the attack on an EPYC 7313 processor based on Zen 3, alongside EPYC 9124 and 9334 processors based on Zen 4. They did not demonstrate the same memory leak on Zen 5 processors.
Beyond confidential virtual machines, the Skyfall researchers also demonstrated that they could recover decrypted microcode updates from all three tested processors.
They further extracted processor register values stored in System Management Mode (SMM) memory, an area normally inaccessible even to the operating system kernel. These additional attacks require privileged software access to the affected machine.
The IEEE Symposium on Security and Privacy programme committee, which accepted the Skyfall research paper, raised concerns about the practical conditions required to carry out the attack.
These include having a processor core without competing activity, repeatedly triggering failed virtual machine startups, performing performance-counter reads at high speed and ensuring the targeted virtual machine does not crash or time out during the process.
Affected AMD Processors and Security Fixes
AMD has assigned separate vulnerability identifiers to the two attacks.
Mior is tracked as CVE-2026-0462 and carries a CVSS 4.0 score of 5.9. Skyfall’s virtual machine register-data exposure is tracked as CVE-2026-40671, with a severity score of 4.9.
AMD identifies SEV-SNP as the affected technology for Mior, excluding the earlier SEV and SEV-ES versions.
The company’s security advisories identify five affected EPYC processor families and outline the corresponding firmware updates.
For Mior, AMD lists SEV firmware versions released on August 12, while the corresponding platform firmware updates are targeted for December 2026.
For Skyfall, platform firmware updates are targeted for November 2026, with microcode updates dated August 12.
The affected processor families include EPYC 7003, 8004, 8005, 9004 and 9005 series.
AMD also identified the Instinct MI300A accelerator as affected by Skyfall, with a patch dated September 21. Security updates for embedded EPYC processors are expected between January and April 2027.
The researchers behind both disclosures said mitigation updates had already been released, although AMD’s advisories do not clearly establish whether the firmware and microcode updates can be installed independently or must be deployed together through BIOS updates.
The advisories also do not provide alternative workarounds for affected systems.
Server manufacturer Supermicro has published a list of BIOS updates addressing vulnerabilities across dozens of its H12, H13 and H14 motherboard models. The company also says affected systems require Baseboard Management Controller (BMC) firmware updates, while testing of some products remains ongoing.
Because the fixes involve firmware installed on the physical host, customers renting confidential virtual machines will depend on their cloud providers to deploy the necessary updates.
AMD’s Mior advisory also provides Trusted Computing Base (TCB) values for SEV-SNP attestation alongside the fixed firmware versions. However, it does not explain how virtual machine owners should verify those values.
Neither AMD advisory confirms whether the vulnerabilities have been exploited in actual cyberattacks. Both research teams have released technical materials and proof-of-concept code.
The Mior demonstration code, however, requires a BIOS configuration that leaves the processor’s memory-routing settings unlocked.
Additional EDK II Vulnerability
A related security issue affecting EDK II firmware code has also been assigned a separate vulnerability identifier.
The issue, tracked as CVE-2026-18448, concerns Skyfall’s exposure of sensitive information stored in System Management Mode memory. It carries a CVSS score of 5.6.
AMD’s advisory attributes the weakness to missing mitigations against Transient Scheduler Attacks in EDK II firmware code.
The affected systems include several EPYC and Ryzen processor models, including embedded versions. AMD recommends applying its previously released TSA mitigations alongside an EDK II code update when available.
For the five EPYC server families identified in the Mior and Skyfall disclosures, the relevant EDK II patch is targeted for November 2026. Firmware updates addressing the issue are already available for some other processors, including EPYC 4004 and 4005 models.
However, the EDK II project’s own advisory had not listed a patched version as of the publication date.
The disclosures highlight continuing security challenges in confidential computing, where hardware-based encryption and isolation are designed to protect sensitive workloads even when they run on infrastructure controlled by third parties.