Yaamlabs

Vulnerabilities

Skyfall and Mior break AMD confidential VMs

Two attacks from ETH Zurich and Google weaken AMD SEV-SNP, the tech that hides cloud VMs from the host. What they do and what tenants can actually fix.

Two attacks disclosed in mid-November 2026 weaken AMD SEV-SNP, the technology meant to stop a cloud provider from reading or changing the virtual machines it hosts. SEV-SNP stands for Secure Encrypted Virtualization with Secure Nested Paging; it keeps a guest virtual machine's (VM's) memory encrypted with a key the host never sees, and lets the guest prove, through a signed measurement, that it is running on genuine, patched AMD hardware before any secret is released to it. Researchers from ETH Zurich, with a collaborator from Google, showed two separate ways to break that promise. If you run confidential VMs or confidential artificial-intelligence (AI) workloads on AMD EPYC processors in the cloud, this concerns you.

The attacker here is not a stranger on the internet. It is whoever controls the machine your confidential VM sits on: a malicious cloud operator, a compromised hypervisor, a rogue provider insider, or someone with physical and firmware access to the server. Defending the guest from that party is the whole reason SEV-SNP exists, so these flaws matter more than their modest severity scores suggest. Both fixes live on the host, which means most readers, who are tenants, cannot apply them. A tenant's real lever is attestation policy and pressure on the provider, which the sections below lay out.

How it works

A confidential VM keeps its memory encrypted, and a hardware structure called the Reverse Map Table (RMP) records which physical memory page belongs to which guest, so the host cannot quietly swap or alias a page underneath the guest. Before an application trusts such a VM with secrets, it asks for remote attestation: the AMD Secure Processor signs a report that contains a measurement of the VM's starting state and the firmware version, called the Trusted Computing Base (TCB) level, and a verifier checks that signature and those values. If the measurement or the TCB level is wrong, the verifier withholds the secret. Both attacks aim at this machinery.

Skyfall (CVE-2026-40671, AMD bulletin AMD-SB-3061) is information disclosure only. CVE stands for Common Vulnerabilities and Exposures, the public identifier for a flaw. When a VMRUN instruction, the instruction that enters a guest, hits an error and the processor unwinds it, the cleanup is incomplete and leaves guest register state behind, and a transient-execution side channel lets the host read what remains. In testing the researchers recovered a confidential VM's disk-encryption key this way. Skyfall only reads; it cannot change guest memory. It needs control of the host hypervisor, along with precise scheduling and access to performance counters. AMD rates it 4.9 on the CVSS 4.0 scale (CVSS is the Common Vulnerability Scoring System, which runs from 0 to 10).

Mior (CVE-2026-0462, AMD bulletin AMD-SB-3059) is the more damaging of the two. During startup the Secure Processor writes the RMP into memory. Mior redirects those writes to memory on an attacker-controlled PCIe device (PCIe, Peripheral Component Interconnect Express, is the bus that graphics cards and other add-in devices use); the researchers used an NVIDIA graphics card. Part of the genuine RMP, including the entry that protects the table itself, is left uninitialised, so the host can then rewrite ownership records, switch on guest debugging, and read or write confidential VM memory directly. It can also copy a harmless VM's measurement into a malicious one, so the attestation report that should catch the tampering instead vouches for it. Mior needs control of the hypervisor and of the BIOS (Basic Input/Output System, the firmware that configures the server at power-on), as well as the PCIe device. AMD rates it 5.9 on CVSS 4.0.

What the attacks do

Skyfall was demonstrated on Zen 3 (EPYC 7313) and Zen 4 (EPYC 9124 and 9334) processors, and not on Zen 5. Mior was demonstrated on a single Zen 5 chip, an EPYC 9135, where the debugging attack succeeded in all 1,000 trials across the test runs, taking about 1.7 milliseconds each, and worked whether or not the optional ciphertext-hiding feature was on. The researchers believe Zen 3 and Zen 4 are also exposed to Mior but did not test them. Between the two papers that covers Zen 3 through Zen 5, but no single attack has been shown across the whole range, and the sources disagree on exactly which generation each flaw reaches.

A few other gaps are worth stating plainly. AMD-SB-3061 appears under different titles across write-ups, with AMD's own page calling it an incomplete-cleanup issue, which makes it easy to miss when searching. One analysis notes that AMD-SB-3061 lists no TCB values for Skyfall, so a verifier has no published threshold to reject reports against; the Mior bulletin, AMD-SB-3059, does list a minimum SNP TCB level per EPYC family. The advisories also do not say whether the microcode and firmware updates work on their own or only once the BIOS update is in place. And neither AMD nor any cloud provider has reported either flaw being used in a real attack.

What to do

If you run AMD confidential VMs in the cloud, you cannot patch the host, so attestation is your control:

  • Update attestation verification to require the fixed firmware and TCB level, and reject any report below it. For Mior this is workable, because AMD-SB-3059 publishes the minimum SNP TCB level per EPYC family. For Skyfall there is no published TCB value yet, so ask your provider how a Skyfall-patched platform will show up in an attestation report.
  • Confirm with your provider that it has applied AMD's SEV firmware, the microcode, and the server maker's BIOS updates.
  • Treat attestation as the primary control rather than a formality. A verifier that accepts any signed report gives up the protection entirely.
  • Until a platform reaches the fixed TCB level, reconsider what secrets you place in a confidential VM on it.

If you run AMD EPYC confidential computing on your own hardware:

  • Apply the SEV firmware and microcode updates dated 12 August 2026, then the platform firmware and BIOS updates from your server vendor as they ship. AMD targets November 2026 for the Skyfall platform firmware and December 2026 for Mior. The Instinct MI300A accelerator fix is dated 21 September 2026, and embedded EPYC parts are not scheduled until January to April 2027.
  • Track AMD-SB-3059 and AMD-SB-3061 for revised dates, since the BIOS dates are targets rather than confirmed releases.
  • Protect BIOS integrity, because Mior depends on BIOS control: use verified or secure boot and restrict who can change firmware settings. Some server vendors also require a baseboard management controller (BMC) firmware update alongside the BIOS.

The wider lesson

Confidential computing holds its guarantees only when the trusted computing base is current and attestation is genuinely checked. A confidential VM running on firmware below the fixed TCB level, or verified by a service that waves through whatever the host sends, offers less than it appears to. Skyfall and Mior are a reason to keep reviewing the TCB level you accept and the attestation policy that enforces it, rather than setting them once and leaving them.

We help teams running confidential and AI workloads get this part right, from attestation policies that reject stale TCB levels to how secrets reach a guest in the first place. See our work on cloud and Kubernetes security and AI-native systems, or safeguarding and hardening for firmware and boot integrity on servers you run yourself. Open the chat and Yaali, our AI agent, will pass your question to the engineer who would do the work.


Sources: CSA research note on Mior and Skyfall, AMD-SB-3061 (Skyfall), AMD-SB-3059 (Mior), Simply Secure Group write-up, Spartech Software summary, CASRAI overview.

Back to the blog, or read this post on the full site.