Two attacks published this week allow a malicious host to read data inside AMD SEV-SNP-protected virtual machines, and one of them goes further.
The attack called Mior can also modify a VM's memory and fake the signed report confirming which software it runs. That second capability matters more than the first, because the entire premise of confidential computing is that you do not have to trust the host. The attestation report is how you verify that premise.
Both attacks need control of the server's hypervisor. Mior also needs control of the BIOS. Researchers demonstrated both in lab tests on AMD EPYC server processors.
What SEV-SNP Is Supposed to Do
SEV-SNP brings confidential computing to AMD EPYC server processors. It uses hardware-enforced memory encryption and access control to protect a VM, even from the host's hypervisor.
The threat model is explicit. You are running on hardware someone else controls, and you want the guarantee that the operator cannot read your memory or tamper with your workload.
Mior attacks that guarantee during startup.
How Mior Works
During VM startup, a security co-processor on the chip called the Platform Security Processor fills in a table called the Reverse Map Table. The RMP records who owns each page of memory and what they are permitted to do with it.
It is the core enforcement structure. Corrupt it, and the hardware protections stop meaning anything.
The ETH Zurich team behind Mior redirected the PSP's writes to the memory of a graphics card it controlled, rather than to main memory.
The choice of a graphics card is deliberate. The team believes the PSP reads values back to verify them, and the card could return whatever the verification expected to see. Startup then completed without errors.
But part of the real RMP, including the entry that protects the table itself, stayed uninitialized.
That gap let the hypervisor rewrite the RMP's protections. The team then enabled debugging on a confidential VM that had debugging disabled, and used the PSP's own debug commands to read and write the VM's memory.
It worked in all 1,000 attempts across both test runs. That is not a probabilistic result.
The Attestation Forgery
SEV-SNP measures a VM's software image when the VM starts. The PSP signs attestation reports that include that measurement.
In a second test, the team copied a harmless VM's measurement into a malicious VM. The malicious VM's attestation reports then described the harmless image instead of what was actually running.
That is the piece with the widest implications. Confidential computing customers use attestation to verify what they are running on. If the report can be made to describe software that is not there, the verification becomes decorative.
What Mior Requires
The attack needs control of the hypervisor and the BIOS, also called UEFI, plus a PCIe device such as a graphics card.
For the tests, the team modified the BIOS to leave the chip's memory-routing settings unlocked. The Mior paper states the attack operates within AMD's threat model for SEV-SNP.
The researchers' position is direct. AMD explicitly distrusts the UEFI, thus an attacker can modify the UEFI.
That is a fair reading. The BIOS was already outside the trusted computing base for SEV-SNP. Mior demonstrates what happens when that assumption is tested.
What Mior Does Not Break
Ordinary VMs already trust their host and its firmware, so Mior gives an attacker nothing new against them, the team says.
The attack is specifically a problem for the confidential computing use case, where the host is not supposed to be trusted.
Testing Scope
The team ran Mior on one EPYC 9135, a Zen 5 chip. It considers Zen 3 and Zen 4 chips potentially affected but did not test them, because it had no setup that allowed it to change their BIOS.
The debugging and attestation tests both worked with ciphertext hiding turned on. That is an optional SEV-SNP feature that stops the hypervisor from even reading encrypted VM memory.
Relationship to Earlier Attacks
Two earlier attacks by some of the same researchers, Fabricked and RMPocalypse, also corrupted the RMP as SEV-SNP started up.
Mior is a separate flaw, the team says, unrelated to Fabricked and to the fix for it. It also worked on older firmware from before that fix.
How Skyfall Works
Skyfall comes from a team at ETH Zurich and Google. It builds on Transient Scheduler Attacks, a group of flaws in AMD processors that can leak data through instruction timing. TSA was first disclosed on July 8, 2025.
The mechanism is specific. When a hypervisor intentionally causes a VM to fail to start, the processor has already cached part of the VM's saved state. That data is left behind without the usual cleanup, and the hypervisor can read it through TSA.
In the lab, the team recovered a VM's saved register values and a disk-encryption key.
For the key, it changed the VM's kernel to pause at the moment the key sat in registers. In a real attack, the team says a hypervisor can instead use a technique called SEV-Step to step through the VM until the key reaches the registers.
That is the difference between a demonstration and a practical attack, and the team is honest about which one it built.
The Limits of Skyfall
Skyfall can read VM memory, but only three fixed cache lines in a targeted page, at offsets 0x0c0, 0x380, and 0x3c0. The rest of the page stays out of reach.
It needs a malicious host or hypervisor, but no tampering with the hardware.
The team tested it on an EPYC 7313, a Zen 3 chip, and on EPYC 9124 and 9334 chips, which are Zen 4. It did not show the leak on Zen 5 chips.
Beyond VMs
The Skyfall team also read decrypted microcode updates on all three test chips. And it recovered register values stored in System Management Mode memory, which even the operating system kernel normally cannot read.
Both of those attacks need privileged software on the machine. They are not available to a tenant.
The Review Committee's Critique
The program committee of the IEEE Symposium on Security and Privacy, which accepted the Skyfall paper, said it overlooks what a realistic attack on VMs would need.
That includes a processor core free of other activity, failing VM starts and performance-counter reads at high speed, and a VM that does not crash or time out during the attack.
That critique is worth reading alongside the paper. It is the reviewers saying the demonstrated attack is real but the operational version is harder than it looks.
The Severity Scores
AMD tracks Mior as CVE-2026-0462 and rates it 5.9 on the CVSS 4.0 scale. It tracks Skyfall's VM leak as CVE-2026-40671, rated 4.9.
Those numbers are low for attacks that break confidential computing. The reason is the precondition, not the impact. Both require a malicious hypervisor, and the CVSS framework weights that heavily. If you are a cloud tenant, the hypervisor is exactly what you do not control.
For Mior specifically, AMD lists only SEV-SNP as affected, not the earlier SEV and SEV-ES versions.
Affected Processors and Fixes
AMD lists the same five EPYC server lines for both flaws.
|
EPYC Series |
Mior Platform Firmware (Target) |
Mior SEV Firmware (Aug 12) |
Skyfall Platform Firmware (Target) |
|
7003 (Milan) |
MilanPI-SP3_1.0.0.L (December) |
1.3A.03 |
MilanPI-SP3_1.0.0.L (November) |
|
8004 (Genoa) |
GenoaPI_1.0.0.K (December) |
1.3A.4 |
Genoa_100J Patch 3 (November) |
|
8005 (Turin) |
TurinPI_1.0.0.C (December) |
1.3A.0A |
TurinPI_1.0.0.C (November) |
|
9004 (Genoa) |
GenoaPI_1.0.0.K (December) |
1.3A.4 |
Genoa_100J Patch 3 (November) |
|
9005 (Turin) |
TurinPI_1.0.0.C (December) |
1.3A.0A |
TurinPI_1.0.0.C (November) |
Three additional details from the advisories.
The two advisories give the same TurinPI_1.0.0.C package different target months, which is worth verifying with your server maker before planning a window.
Microcode for Skyfall is dated August 12 for each of these lines. The Instinct MI300A accelerator has a Skyfall patch dated September 21, and fixes for embedded EPYC models are due between January and April 2027.
AMD says it plans to give server makers platform firmware in November for Skyfall and in December for Mior.
The Firmware Supply Chain Problem
Here is the operational reality that matters most.
The fixes are implemented in the host's firmware. Customers who rent confidential VMs depend entirely on their cloud provider to install them.
There is no patch a tenant can apply. There is no configuration a tenant can change. The entire remediation is upstream of you.
For organizations running their own EPYC hardware, the fix path runs through the server maker. Supermicro has published a BIOS fix list covering dozens of its H12, H13, and H14 motherboards. The same pages state the boards also need a BMC firmware update, and that the company is still testing affected products.
The advisory covering Mior also lists TCB values for SNP attestation next to each fixed firmware version. It does not say how VM owners should check them.
That is the gap worth flagging. If attestation TCB values are how a tenant verifies the fix is in place, and nobody has published a verification procedure, the verification is theoretical.
The EDK II Piece
The EDK II project assigned Skyfall's SMM leak CVE-2026-18448 and rates it 5.6.
AMD's EDK II bulletin traces it to missing TSA mitigations in EDK II firmware code. The bulletin lists EPYC and Ryzen chips, including embedded models, and recommends the company's earlier TSA mitigations plus an EDK II patch when one is available.
For the five EPYC server lines above, the bulletin lists only that patch, due in November. Platform firmware with the fix is already available for other chips, including the EPYC 4004 and 4005 models.
The EDK II project's own advisory lists no patched version as of publication.
Exploitation Status
Neither advisory says whether the flaws have been used in attacks.
Both teams have published code. The Mior proof-of-concept needs a BIOS patched to leave the memory-routing settings unlocked, which is not a condition an attacker can arrange remotely.
What Defenders Should Do
- If you rent confidential VMs, ask your provider when the firmware lands. That is the entire remediation path. Get a date, get a version, and ask how you can verify it.
- If you run your own EPYC hardware, go to your server maker. AMD is supplying platform firmware to OEMs in November and December. Supermicro has already published BIOS fix lists, along with a note that BMC firmware also needs updating.
- Check the Turin month discrepancy. The two advisories list different target months for the same TurinPI package. Confirm with your vendor rather than assuming either date.
- Do not assume the SEV firmware alone is sufficient. The advisories do not state whether the August 12 firmware works standalone or only as part of the BIOS update. Plan for the full stack.
- Track the EDK II patch. It covers the SMM leak separately, and it is due in November for the EPYC server lines.
- Reassess what attestation proves. Mior demonstrates that a malicious host with BIOS control can make a report describe software that is not running. If your threat model includes a hostile cloud provider, attestation is now a weaker guarantee than it was.
- Note the ciphertext-hiding caveat. Both tests worked with that optional feature enabled. Enabling it does not mitigate Mior.
- Watch for the TCB verification guidance. If AMD or your provider publishes a method for checking attestation TCB values against fixed firmware, use it. Until then, treat the fix as unverified.
FAQ
What is Mior?
An attack on AMD SEV-SNP that corrupts the Reverse Map Table during VM startup, letting a hypervisor with BIOS control read and modify confidential VM memory and forge attestation reports.
What is Skyfall?
An attack that builds on Transient Scheduler Attacks. When a hypervisor intentionally causes a VM to fail to start, cached state is left behind without cleanup, and the hypervisor can read it. In the lab, it recovered register values and a disk-encryption key.
Do these break confidential computing?
Mior does, under its stated preconditions. It requires a malicious hypervisor plus BIOS control plus a PCIe device. The attestation forgery is the most significant part, because attestation is how a tenant verifies what is running.
Can a cloud tenant exploit this?
No. Both attacks require control of the hypervisor. A tenant does not have that. The relevance to tenants is that they are the ones who need the host patched.
Which processors are affected?
AMD lists five EPYC server lines for both flaws: 7003 (Milan), 8004 (Genoa), 8005 (Turin), 9004 (Genoa), and 9005 (Turin).
What are the CVE numbers and scores?
Mior is CVE-2026-0462, rated 5.9. Skyfall's VM leak is CVE-2026-40671, rated 4.9. Skyfall's SMM leak is CVE-2026-18448, rated 5.6 by the EDK II project.
Why are the scores so low for attacks on confidential computing?
Because both require a malicious hypervisor, and CVSS weights that precondition heavily. The impact is high, but the attack path requires control the framework treats as significant.
When are fixes available?
AMD lists SEV firmware for Mior and microcode for Skyfall, both dated August 12. Platform firmware goes to server makers in November for Skyfall and December for Mior. Embedded EPYC fixes are due between January and April 2027.
Are the fixes enough on their own?
The advisories do not say whether the August 12 firmware and microcode work standalone or only as part of the BIOS update. They also list no workaround.
How do I verify the fix?
The Mior advisory lists TCB values for SNP attestation next to each fixed firmware version, but does not say how VM owners should check them. That is a gap.
Does enabling ciphertext hiding help?
No. Both the debugging and attestation tests worked with ciphertext hiding turned on.
Has either attack been used in the wild?
Neither advisory reports exploitation. Both teams published code, and the Mior proof of concept requires a BIOS patched to leave memory-routing settings unlocked.
What did the IEEE reviewers say about Skyfall?
That the paper overlooks what a realistic VM attack would need, including a free processor core, high-speed performance-counter reads, and a VM that does not crash or time out.
Is Mior related to Fabricked or RMPocalypse?
No. The researchers say it is a separate flaw, unrelated to Fabricked and to the fix for it. It also worked on older firmware from before that fix.