Security & privacy

Mior and Skyfall attacks breach AMD SEV-SNP confidential VMs

New Mior and Skyfall attacks allow malicious hosts to read secrets from AMD SEV-SNP protected VMs, requiring firmware updates for mitigation.

Illustration of a server rack with a broken security shield
Illustration generated for this article

Two new attack vectors, Mior and Skyfall, have been disclosed that allow a malicious cloud host to extract secrets from virtual machines protected by AMD’s SEV-SNP confidential computing technology. Researchers from ETH Zurich and Google demonstrated these flaws in laboratory settings on AMD EPYC server processors, revealing weaknesses in how the hardware handles memory protection and state cleanup during startup failures. AMD has released microcode and firmware patches dated August 12, with broader platform firmware updates scheduled for distribution through server manufacturers in November and December 2026.

What happened

The Mior attack targets the initialization phase of SEV-SNP, allowing an attacker with control over the hypervisor and the server’s BIOS to bypass memory encryption. By redirecting writes from the Platform Security Processor (PSP) to a controlled PCIe device, such as a graphics card, the researchers left parts of the Reverse Map Table (RMP) uninitialized. This table is critical for defining memory ownership and access rights. With the RMP compromised, the hypervisor could rewrite protections, enable debugging on a locked-down VM, and read or write its memory using the PSP’s own debug commands. The attack succeeded in all 1,000 attempts across two test runs and also allowed the attackers to forge attestation reports by copying measurements from a benign VM to a malicious one.

Skyfall, developed by a joint team from ETH Zurich and Google, exploits a different mechanism related to Transient Scheduler Attacks (TSA). When a hypervisor intentionally causes a VM to fail during startup, the processor caches part of the VM’s saved state but fails to clean it up properly. The hypervisor can then use timing side-channels to read this leftover data from the cache. In tests, researchers recovered saved register values and even a disk-encryption key by pausing the VM kernel at the precise moment the key was loaded into registers. Unlike Mior, Skyfall does not require BIOS tampering, only a malicious hypervisor. It was successfully demonstrated on Zen 3 and Zen 4 EPYC chips but did not show the same leakage on Zen 5 hardware.

Both attacks operate within AMD’s defined threat model for SEV-SNP, which explicitly distrusts the UEFI firmware and assumes the hypervisor may be hostile. However, they challenge the practical security guarantees expected by users of confidential computing. AMD has assigned CVE-2026-0462 to Mior with a CVSS 4.0 score of 5.9, and CVE-2026-40671 to Skyfall with a score of 4.9. The company states that fixes are available in SEV firmware and microcode, but full mitigation requires platform firmware updates from server makers like Supermicro, which are rolling out in phases through late 2026 and early 2027.

How it works

SEV-SNP protects VMs by encrypting memory and using the RMP to enforce access controls, ensuring that even the host hypervisor cannot read or modify VM data. Mior subverts this by interfering with the PSP during boot. The PSP normally populates the RMP in main memory, but Mior tricks it into writing to a PCIe device’s memory instead. Because the PSP reads back these values to verify them, the malicious device can return fake data, leaving the actual RMP in main memory partially empty. This gap allows the hypervisor to alter memory permissions and inject debug commands, effectively unlocking the VM from the inside. The attack also exploits the attestation process, where the PSP signs a report of the VM’s software state. By swapping measurements, attackers can make a malicious VM appear legitimate.

Skyfall leverages the fact that processors cache data for performance. When a VM start-up is aborted, the CPU may retain fragments of the VM’s state in its cache lines. Skyfall uses TSA techniques to measure instruction timing and infer the content of these cache lines. Specifically, it can read three fixed offsets in a targeted memory page. While this provides limited memory access, it is sufficient to extract high-value targets like encryption keys if the attacker can control the VM’s execution flow to place secrets in those specific cache locations. The attack also extends to reading decrypted microcode updates and System Management Mode (SMM) memory, which is typically isolated from the operating system.

Key details

  • Mior requires control of the hypervisor, BIOS, and a PCIe device, while Skyfall only requires a malicious hypervisor.
  • Mior was tested on an EPYC 9135 (Zen 5), while Skyfall was demonstrated on EPYC 7313 (Zen 3) and EPYC 9124/9334 (Zen 4).
  • AMD rates Mior as CVE-2026-0462 (5.9 CVSS) and Skyfall as CVE-2026-40671 (4.9 CVSS).
  • Fixes include SEV firmware and microcode dated August 12, with platform firmware updates arriving November–December 2026.
  • Skyfall can read three specific cache lines per page and has been used to extract disk-encryption keys in lab tests.
  • Neither advisory confirms active exploitation in the wild, but proof-of-concept code has been published by the research teams.

Why it matters

For engineers building products on confidential computing, these findings highlight that hardware-enforced isolation is not absolute. SEV-SNP is often chosen for multi-tenant cloud environments where the provider is untrusted, but both attacks assume a malicious host. This means that organizations relying on SEV-SNP for regulatory compliance or data sovereignty must verify that their cloud providers are applying firmware updates promptly. The dependency on BIOS and platform firmware updates introduces a lag between patch release and actual protection, leaving windows of vulnerability.

The distinction between the two attacks also matters for risk assessment. Mior is more complex to execute, requiring physical or deep firmware access, whereas Skyfall is a pure software-side channel attack that could be more easily deployed by a compromised hypervisor. However, Skyfall’s impact is currently limited to specific cache lines and older architectures. Developers should review their attestation workflows, as Mior demonstrates that signed reports can be forged if the underlying memory protections are bypassed. Trusting attestation alone is insufficient if the initialization process itself is vulnerable.

What you can do

  • Contact your cloud provider or server vendor to confirm the schedule for BIOS and platform firmware updates addressing CVE-2026-0462 and CVE-2026-40671.
  • Verify that your SEV-SNP attestation checks include the latest Trusted Computing Base (TCB) values provided by AMD in their security advisories.
  • Monitor for microcode updates from your hardware vendor, as these are prerequisites for the broader platform firmware fixes.
  • If you manage bare-metal servers, prioritize updating BMC and BIOS firmware, and ensure you are running the latest SEV firmware versions listed by AMD.
  • Review your VM startup procedures to minimize the window where failed starts might leave sensitive data in cache, though this is largely a mitigation for Skyfall on affected Zen 3/4 chips.
  • Stay informed about EDK II patches, as Skyfall also affects SMM memory and microcode handling in the firmware layer.
  • #amd sev-snp
  • #confidential computing
  • #cloud security
  • #hardware vulnerabilities
  • #firmware security

Tools from the Bytechap store

Keep reading

All stories