The Mior and Skyfall Confidential VM Attacks Break AMD SEV-SNP's Core Promise

The Mior and Skyfall confidential VM attacks, disclosed November 13, 2026 by researchers at ETH Zurich and Google, demonstrate that a malicious cloud host can read, and in Mior's case rewrite, the memory of a confidential virtual machine that AMD SEV-SNP is specifically designed to hide from that same host. According to The Hacker News, Mior worked in every single one of 1,000 test trials.

Jump to why tenants can't fix this themselves if you want the part that matters most for anyone renting confidential compute.

What SEV-SNP Was Supposed to Guarantee

AMD's Secure Encrypted Virtualization-Secure Nested Paging is the hardware foundation underneath "confidential computing": memory encryption and access control built into AMD EPYC processors specifically so that even a host with full hypervisor and root access cannot read or alter guest memory. A signed attestation report is supposed to prove to a remote party, before they trust the machine with sensitive data, that the VM is running genuine, unmodified software on genuine AMD hardware. That attestation report is the entire basis on which someone agrees to send secrets to a server they don't control.

How Mior Breaks It: Turning a Graphics Card Into Memory

CASRAI's technical writeup describes Mior's mechanism as a boot-time trick against the Reverse Map Table, the structure AMD's Platform Security Processor uses to track which memory pages belong to which guest. An attacker controlling the hypervisor and the BIOS manipulates memory-mapped I/O routing during boot, redirecting the PSP's protective writes to a secondary PCIe device, such as an ordinary graphics card, instead of system memory.

Anonymoushackers.net's reporting calls this the "Device-as-Memory" technique: because the PCIe device returns responses that look correct during verification, the boot process completes successfully while the real Reverse Map Table entries, including the one meant to protect the table itself, are left uninitialized.

With that protection gone, the hypervisor can enable debugging commands on a running confidential VM that was supposed to have debugging permanently disabled, then freely read and write its memory. Osintsights' detailed account reports the researchers also copied a benign VM's measured boot image into a malicious VM, so the malicious VM's signed attestation reports falsely described the harmless software, defeating the attestation guarantee directly rather than just bypassing memory encryption.

How Skyfall Leaks Secrets Without Touching the BIOS

Skyfall takes a lighter-weight path. It builds on Transient Scheduler Attacks, a CPU timing side-channel class first disclosed in July 2025, and targets the moment a VM's internal register state gets cached during a forced, incomplete startup. Privacyneedle's reporting explains that when a hypervisor deliberately causes a confidential VM to fail to start, the processor caches part of that VM's saved state without performing its usual cleanup.

A malicious host on the same physical core can then recover that cached data through timing analysis, and the research team used this path to recover an actual disk-encryption key during testing, per The Hacker News.

Skyfall requires no BIOS modification and no additional hardware, only hypervisor control, which makes it the lower-effort of the two attacks even though its blast radius is narrower than Mior's full read-write access.

Why Tenants Can't Fix This Themselves

Both flaws live entirely in host-side firmware and platform BIOS, not in anything a tenant renting a confidential VM controls. SparTech's reporting confirms AMD shipped SEV firmware and microcode updates dated August 12, 2026, but the platform-initialization firmware that server manufacturers need to actually ship the fix wasn't scheduled until November 2026 for Skyfall and December 2026 for Mior, delivered through individual OEM BIOS updates from vendors like Dell, HPE, and Supermicro.

Until a specific host receives that update, nothing a customer does inside their own VM closes the gap, because the vulnerability sits below the layer the customer can see or touch.

Neither AMD's advisories nor any of the reporting we reviewed claims these attacks have been used outside a lab setting. Both research teams published proof-of-concept code, and Mior's specifically requires a BIOS left in an unpatched, memory-routing-unlocked configuration to function, which narrows but doesn't eliminate real-world exploitability on hosts that haven't yet received the fix.

What This Means If You Rely on Confidential Computing

Our verdict: this doesn't mean AMD SEV-SNP is broken as a concept, both attacks require the one thing SEV-SNP's threat model already assumes an attacker might have, hypervisor control, and in Mior's case BIOS control too. But it does mean the specific guarantee customers were sold, "even your cloud provider's root access can't reach your data," has a real, demonstrated, 1,000-for-1,000 counterexample right now, on hardware that hasn't received the fix yet.

• Ask your confidential-compute provider directly for firmware and BIOS update timelines, not just a general assurance that "patches are being applied." The fix ships per-host, through vendor BIOS updates, not as a single coordinated rollout.
• If your verification process accepts any validly signed attestation report, tighten it to reject reports below AMD's fixed TCB (trusted computing base) version. An unpatched host can still produce a report that passes a less specific check.
• Treat attestation as a point-in-time claim about firmware state, not a permanent guarantee. A host that was trustworthy last month is only still trustworthy if it actually received the BIOS update in between, and neither advisory describes a way for a VM owner to verify that independently without checking the TCB value themselves.

FAQ

What are the Mior and Skyfall AMD SEV-SNP attacks?

Mior (CVE-2026-0462) and Skyfall (CVE-2026-40671) are two separate attacks, disclosed November 13, 2026 by ETH Zurich and Google researchers, that let a malicious cloud host bypass AMD SEV-SNP's confidential-VM protections. Mior gives an attacker controlling the hypervisor and BIOS full read-write access to a guest's encrypted memory and lets them forge attestation reports; Skyfall lets a host read cached VM register data, including encryption keys, through a timing side channel, without needing BIOS access.

Can I patch my own confidential VM against Mior and Skyfall?

No. Both vulnerabilities exist in host-level firmware and platform BIOS, components the tenant renting a confidential VM does not control. Fixes must be applied by the physical server's owner or operator, typically a cloud provider, through AMD's firmware updates and the corresponding server manufacturer's BIOS release, which AMD scheduled for November 2026 (Skyfall) and December 2026 (Mior).

Have Mior or Skyfall been exploited outside of a research lab?

No public reporting we reviewed, including AMD's own security advisories, describes in-the-wild exploitation of either vulnerability. Both were demonstrated by research teams in controlled lab environments, and both teams have published proof-of-concept code, which narrows the question to how quickly affected hosts receive their firmware updates.

What We Checked Before Publishing

We cross-referenced the technical mechanism for both attacks across CASRAI's, Anonymoushackers.net's, and Osintsights' independent writeups, confirmed CVE identifiers and CVSS scores against AMD's own advisory numbers as cited in multiple outlets, and verified the firmware patch timeline through SparTech's and The Hacker News' reporting. We did not find independent confirmation of in-the-wild exploitation beyond the researchers' own lab demonstrations, and we've reported that distinction rather than implying active attacks are underway.

I'm Rohit, founder of Krova Cloud: Krova's Cubes don't rely on AMD SEV-SNP's memory-encryption-plus-attestation model at all; isolation comes from a per-Cube jailer sandbox and Firecracker microVMs, a different trust boundary with its own tradeoffs, not an immunity claim.

What this research actually demonstrates, that a hardware security guarantee is only as good as the firmware underneath it actually being current, applies to any isolation architecture, including ours. That's exactly why we'd rather tell you what this specific attack doesn't touch than let a marketing page imply nothing ever could. More debugging stories at debugly.dev.

Structured Data

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "BlogPosting",
      "headline": "The Mior and Skyfall Confidential VM Attacks Break AMD SEV-SNP's Core Promise",
      "datePublished": "2026-11-15",
      "dateModified": "2026-11-15",
      "author": {
        "@type": "Person",
        "name": "Rohit"
      },
      "about": ["Mior attack", "Skyfall attack", "AMD SEV-SNP", "confidential computing security"]
    },
    {
      "@type": "FAQPage",
      "mainEntity": [
        {
          "@type": "Question",
          "name": "What are the Mior and Skyfall AMD SEV-SNP attacks?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Mior (CVE-2026-0462) and Skyfall (CVE-2026-40671) are two separate attacks, disclosed November 13, 2026 by ETH Zurich and Google researchers, that let a malicious cloud host bypass AMD SEV-SNP's confidential-VM protections. Mior gives an attacker controlling the hypervisor and BIOS full read-write access to a guest's encrypted memory and lets them forge attestation reports; Skyfall lets a host read cached VM register data, including encryption keys, through a timing side channel, without needing BIOS access."
          }
        },
        {
          "@type": "Question",
          "name": "Can I patch my own confidential VM against Mior and Skyfall?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "No. Both vulnerabilities exist in host-level firmware and platform BIOS, components the tenant renting a confidential VM does not control. Fixes must be applied by the physical server's owner or operator, typically a cloud provider, through AMD's firmware updates and the corresponding server manufacturer's BIOS release, which AMD scheduled for November 2026 (Skyfall) and December 2026 (Mior)."
          }
        },
        {
          "@type": "Question",
          "name": "Have Mior or Skyfall been exploited outside of a research lab?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "No public reporting reviewed, including AMD's own security advisories, describes in-the-wild exploitation of either vulnerability. Both were demonstrated by research teams in controlled lab environments, and both teams have published proof-of-concept code, which narrows the question to how quickly affected hosts receive their firmware updates."
          }
        }
      ]
    }
  ]
}