Mior and Skyfall: new AMD SEV-SNP attacks let a rogue cloud host read confidential VMs — and fake attestation
ETH Zurich and Google break AMD SEV-SNP twice: Mior forges attestation, Skyfall leaks VM secrets. What it means for "confidential AI" and what to check.
On November 10, 2026 AMD published four security bulletins about SEV-SNP, the processor feature that keeps a confidential virtual machine's memory hidden from the cloud host that runs it. Two describe new academic attacks from ETH Zurich and Google, Mior and Skyfall, which let a malicious host read or change data inside those VMs, and in Mior's case fake the attestation report that tells customers the VM can be trusted. Neither attack has been reported in the wild and both need full control of the server, but that is exactly the attacker confidential computing is built to stop, so anyone selling or buying "confidential AI" on AMD hardware should check that the fixes are installed and that attestation actually checks for them.
The short version
Most cloud customers trust their provider: the provider's hypervisor can see everything inside an ordinary virtual machine. Confidential computing removes that trust. With AMD's SEV-SNP (Secure Encrypted Virtualization – Secure Nested Paging), the processor encrypts the VM's memory, keeps a hardware table of which memory pages belong to which VM, and signs an "attestation report" that a customer can check before sending secrets. This is the foundation of many "confidential AI" offerings. Microsoft's Azure confidential GPU VMs, for example, combine AMD SEV-SNP processors with NVIDIA H100 GPUs, and Microsoft describes the trusted environment as spanning "confidential VM on the CPU and attached GPU."
The two new attacks break that promise from different directions:
- Mior (paper, ACM CCS 2026) tricks AMD's security processor into writing part of its protection table to an attacker-controlled PCIe card (the researchers used a consumer NVIDIA graphics card) instead of memory. The table is left partly unset, and the host can then switch on debugging for a VM that had it disabled, read and write its memory, and make a malicious VM present the attestation identity of a harmless one. AMD tracks it as CVE-2026-0462 in AMD-SB-3059.
- Skyfall (paper, accepted to IEEE S&P 2027) is a side-channel attack. When the host deliberately makes a confidential VM fail to start, the processor has already pulled some of the VM's saved register state into its cache and skips the usual cleanup. The researchers read that data through a known class of AMD timing leaks and recovered a test VM's disk-encryption key. AMD tracks it as CVE-2026-40671 in AMD-SB-3061.
AMD has published fixed firmware and microcode for most affected EPYC server processors, with some platform firmware packages and all embedded parts still scheduled. The practical question for most organisations is not "am I hacked?" but "has my provider, or my own server team, installed the updates, and does my attestation check reject hosts that haven't?"
Timeline
- 2025-08 — ETH Zurich reports the earlier Fabricked attack to AMD, according to Tom's Hardware.
- 2026-02-12 — ETH Zurich reports Mior to AMD; embargo agreed until November 10 (Mior paper, ethics section).
- 2026-04-14 — AMD publishes AMD-SB-3034 (CVE-2025-54510, "SEV-SNP Routing Misconfiguration"), the Fabricked fix.
- 2026-04-21 — ETH Zurich and Google report Skyfall's SEV-SNP findings to AMD; follow-up on microcode and SMM impact on 2026-05-28 (Skyfall paper).
- 2026-05-19 — Fabricked goes public (Tom's Hardware), USENIX Security 2026.
- 2026-08-12 — release date AMD lists for the fixed SEV firmware (Mior) and microcode (Skyfall) for the affected EPYC lines.
- 2026-09-21 — release date AMD lists for the Skyfall fix on the Instinct MI300A.
- 2026-11-10 — embargo ends. AMD publishes AMD-SB-3059 (Mior), AMD-SB-3061 (Skyfall), plus AMD-SB-3058 (CVE-2025-54507, an internally found RMP validation flaw) and AMD-SB-3064 (a Rowhammer attack on the same protection table).
- 2026-11-13 — The Hacker News reports both attacks.
- 2026-11 (target) — Skyfall platform firmware packages for EPYC; 2026-12 (target) — Mior platform firmware packages; 2027-01 to 2027-04 — embedded EPYC fixes (AMD bulletins).
What we know
AMD's bulletins. AMD-SB-3059 says a malicious hypervisor "may be able to tamper with" memory-mapped I/O (MMIO) routing and compromise SEV-SNP integrity. CVE-2026-0462 has a CVSS 4.0 score of 5.9 and affects only SEV-SNP, not the older SEV or SEV-ES modes. AMD-SB-3061 describes "incomplete cleanup during VMRUN error handling" that could expose guest register state through a Transient Scheduler Attack (TSA) side channel; CVE-2026-40671 scores 4.9. Both list the EPYC 7003, 8004, 8005, 9004 and 9005 series. Neither bulletin says whether the flaws have been exploited. AMD tells customers to get BIOS updates from their server maker.
Who found them. AMD credits Chris Wech, Philipp Giersfeld, Benedict Schlüter and Shweta Shinde of ETH Zurich for Mior, and Alessandro Vezzu, Benedict Schlüter, Shweta Shinde (ETH Zurich) and Daniel Moghimi (Google) for Skyfall. The Rowhammer report in AMD-SB-3064 comes from researchers at the Chinese Academy of Sciences' Institute of Information Engineering and Tsinghua University. AMD-SB-3058 was found internally.
Mior's results. The paper says its debug-based read/write attack succeeded in all 1,000 runs, averaging about two attempts each because each try has roughly a 50% chance of working. The authors say attestation was forged "with 100% accuracy." In their words: "Our case studies succeed irrespective of whether the ciphertext hiding feature was on or off." That is AMD's optional extra protection for encrypted memory. Testing was on a single Zen 5 EPYC 9135. The authors list Zen 3 and Zen 4 as potentially affected based on the shared routing logic but did not test them. The attack needs hypervisor control, a modified BIOS that leaves the memory fabric unlocked, a host kernel patch and a PCIe device. The project page says the authors assume cloud providers run custom UEFI firmware, and that AMD's threat model distrusts it.
Skyfall's results. The paper says the team recovered a full AES-XTS disk-encryption key "with 100% accuracy," but the moment the key sat in registers was emulated by modifying the guest kernel, not reached by a fully automated attack. Beyond register state, only three fixed cache lines per targeted memory page can be read. On the same chips the team also read decrypted microcode updates and values from System Management Mode memory. The project page says no physical access is needed and that Skyfall "does not break the underlying encryption algorithm"; it reads data after the CPU has decrypted it internally. Testing was on EPYC 7313 (Zen 3) and EPYC 9124 and 9334 (Zen 4). The leak was not shown on Zen 5. The paper includes the S&P program committee's meta-review, which says the attack assumes conditions a real attacker may not get: a quiet isolated core, very frequent failed VM starts and counter reads, and a VM that does not crash or time out.
Fix status, as published. For Mior, AMD lists SEV firmware dated 2026-08-12 and platform firmware packages (MilanPI-SP3_1.0.0.L, GenoaPI_1.0.0.K, TurinPI_1.0.0.C) targeted for December 2026, with new SNP TCB values that attestation can check. For Skyfall, AMD lists microcode dated 2026-08-12 and platform packages targeted for November 2026. Embedded EPYC fixes are scheduled for January–April 2027. Supermicro's security centre lists both bulletins as affecting BIOS on its H11–H14 product lines, without version numbers when we checked on November 16.
Cloud providers. We found no public statement on Mior or Skyfall from AWS, Microsoft Azure or Google Cloud as of November 16. Google's Confidential VM security bulletins page had no entry for either CVE; it did list the earlier Fabricked fix (GCP-2026-019, April 14) as "mitigated by Google" with no customer action.
Technical analysis
Two different holes in the same wall. SEV-SNP rests on three things: memory encryption, the Reverse Map Table (RMP) that records which VM owns each physical page, and attestation signed by AMD's Platform Security Processor (PSP). Mior attacks the RMP's set-up. Skyfall goes around all three by reading plaintext from inside the processor.
Mior: the security processor trusted a map the host could bend. At start-up the PSP builds a memory map from routing registers in AMD's Data Fabric, the interconnect between cores, memory and devices. The paper says the PSP reads those registers with a broadcast read that combines the values from all fabric entry points. A host that configures some entry points differently can hide a malicious range in that combined value, so the PSP believes an address is DRAM when its writes are actually routed to a PCIe device. Because the device stores what the PSP writes and returns it on read-back, the PSP's own sanity check passes. The authors call this primitive "Device-as-Memory." Part of the RMP, including the entry that protects the RMP itself, is never initialised, and from there the hypervisor can rewrite page protections. It is the fourth of what the authors call "Interconnect Corruption Attacks" on SEV-SNP after Fabricked, BreakFAST and STALEUS. The team says Mior is unrelated to the Fabricked fix and works on firmware predating it.
The root cause is a classic one: a trusted component built its view of the world from configuration controlled by the party it is meant to distrust. AMD's earlier fix closed one path, a missing lock check. Mior shows the PSP's memory map could still disagree with where accesses actually go. The authors' proposed defences, such as moving MMIO set-up into the PSP or enforcing one consistent routing configuration across the fabric, are structural. They are not sure all of them can be done in firmware alone.
Skyfall: cleanup skipped on an error path. AMD's mitigation for Transient Scheduler Attacks (first disclosed in July 2025) clears sensitive processor state with the VERW instruction on privilege transitions. The Skyfall authors found paths that skip it. The important one is a VM start (VMRUN) that the host deliberately makes fail: by then the hardware has already loaded part of the VM's saved state into the L1 data cache, but the abort path returns without the usual confidential-exit cleanup. The researchers combined this with a reverse-engineered model of the cache's way predictor and performance counters to read that data. Error and abort paths are often the least-tested code in any system, and in hardware they are rarely part of the mental model of "transitions" that a mitigation must cover.
The AI angle. Neither paper tests AI workloads; both mention large language models only in related work (an earlier Intel TDX side-channel, TDXRay, that leaked LLM prompts). The connection is architectural. In GPU confidential computing, the protected boundary starts in a CPU confidential VM: NVIDIA's confidential containers documentation lists AMD SEV-SNP (Genoa/Milan) and Intel TDX hosts, and Azure's confidential H100 VMs run on 4th-gen EPYC with SEV-SNP.
Our assessment: in such a stack, the CPU-side confidential VM is where prompts, model weights, keys and the code that verifies the GPU all live. If a malicious host can read or rewrite that VM, or forge its attestation, the GPU's protections no longer guarantee what customers think they do. Mior's attestation forgery is the more serious of the two for confidential AI: the whole "verify before you send your prompt" model assumes the report describes the code that will actually handle it.
There is an irony worth noting: Mior used a graphics card as the attacker's memory stand-in. Confidential AI servers are full of powerful PCIe devices that the host configures. That does not make AI servers uniquely vulnerable, because any PCIe device would do, but it is a reminder that the device fabric is part of the attack surface.
Keep the threat in proportion. Both attacks require the attacker to be the host: hypervisor control, and for Mior control of the BIOS too. Ordinary VMs already trust their host, so for them nothing changes, as the Mior authors themselves note. The realistic adversaries are a malicious or compromised cloud operator, a rogue insider, or an attacker who has already taken over a host. That is precisely the case confidential computing sells protection against. Our Phala analysis and the NEAR AI breach showed attackers gaining host-level or management-plane control at "private AI" providers this autumn. Neither incident involved these flaws. But "the attacker owns the host" is not a hypothetical in this market.
What remains unclear
- Are the fixes already deployed? The Mior project page says updated firmware "has been published," and the Skyfall page says AMD "released microcode updates." AMD's own tables show the core SEV firmware and microcode dated August 12, but platform firmware packages targeted for November (Skyfall) and December (Mior). AMD-SB-3061's summary still says AMD "plans to release mitigations." Neither bulletin says whether microcode or SEV firmware alone is enough, or whether the full BIOS package is required.
- Zen 3 and Zen 4 for Mior; Zen 5 for Skyfall. AMD lists all five EPYC lines for both CVEs, but the researchers tested Mior only on Zen 5 and showed Skyfall's leak only on Zen 3 and Zen 4.
- How customers verify. AMD-SB-3059 lists new SNP TCB values, but neither AMD nor the researchers explain how a VM owner should enforce them in attestation policy.
- Cloud providers. No public statement from AWS, Azure or Google that we could find as of November 16. Google's handling of earlier SEV-SNP flaws (patched centrally, "no customer action required") suggests the pattern, but it is not a statement about these CVEs.
- Real-world feasibility of Skyfall. The S&P reviewers' caveats are significant. A production hypervisor would need to stage many failed VM starts on a quiet core without the guest noticing.
- The EDK II angle. The Skyfall site lists an EDK II CVE (CVE-2026-18448) for the System Management Mode leak. The Hacker News reports EDK II's own advisory had no patched version at time of writing; we did not verify that advisory.
Lessons and what to do
For security and infrastructure teams running SEV-SNP themselves:
- Get the BIOS packages from your server vendor that carry the AMD-SB-3059 and AMD-SB-3061 fixes. Check that both the SEV firmware and the microcode versions in AMD's tables are actually loaded, not just that a BIOS update was applied.
- Keep ECC memory enabled. That is AMD's sole recommendation in AMD-SB-3064 for the Rowhammer attack on the RMP.
- Plan for embedded EPYC appliances to stay unpatched into 2027.
For anyone buying confidential VMs or "confidential AI":
- Ask your provider directly whether CVE-2026-0462 and CVE-2026-40671 are mitigated on the hosts serving you, and since when.
- Make attestation policy reject old firmware: require minimum TCB (Trusted Computing Base) versions in your verifier, not just a valid AMD signature. A forged or stale report signed on an unpatched platform is precisely what Mior exploits. Update those minimums every time AMD publishes an SEV bulletin.
- Treat attestation as one control, not the whole story. Keep secrets out of prompts where possible, use short-lived credentials, and log where requests are routed.
For AI builders offering confidential inference:
- Publish which CPU TEE and firmware baseline your service runs on, and how customers can verify it. Four SEV-SNP bulletins in one day is a sign that this baseline will move often.
- Treat the host's device configuration (PCIe, routing, BIOS) as part of your threat model. If you rent from a hyperscaler, get their patch commitments for SEV-SNP bulletins in writing.
For leadership:
- "Confidential" is a property of a specific hardware and firmware version, not of a product name. Hardware isolation keeps being refined through exactly this kind of research. Budget for continuous firmware patching and attestation-policy maintenance, the same way you budget for OS patching.
Related on AI Cyber Brief: Phala "private AI" exposure · NEAR AI SGLang breach · Cloudflare GLM-5.3 GPU memory leak
Sources
- AMD, AMD-SB-3059: AMD Secure Encrypted Virtualization (SEV) Vulnerability (CVE-2026-0462, Mior), November 10, 2026 (primary)
- AMD, AMD-SB-3061: AMD SEV Incomplete Cleanup Vulnerability (CVE-2026-40671, Skyfall), November 10, 2026
- AMD, AMD-SB-3058 (CVE-2025-54507) and AMD-SB-3064 (Rowhammer RMP corruption), November 10, 2026
- AMD, AMD-SB-3034: SEV-SNP Routing Misconfiguration (CVE-2025-54510, Fabricked), April 14, 2026
- Wech, Giersfeld, Schlüter, Shinde (ETH Zurich), Mior paper (ACM CCS 2026) and project page
- Schlüter, Vezzu, Moghimi, Shinde (ETH Zurich, Google), SKYFALL: Microarchitectural Data Sampling on AMD (IEEE S&P 2027) and project page
- Microsoft Learn, Azure confidential GPU options
- NVIDIA, Confidential Containers: supported platforms
- Google Cloud, Confidential VM security bulletins (checked November 16, 2026)
- Supermicro, Security Center (checked November 16, 2026)
Coverage:
- The Hacker News, New Mior and Skyfall Attacks Let Rogue Cloud Hosts Steal Secrets From AMD SEV-SNP VMs, November 13, 2026
- Tom's Hardware, Fabricked flaw lets malicious cloud hosts silently read confidential VM memory, May 19, 2026