CybersecurityNovember 14, 20265 min read

New Mior and Skyfall attacks break a key promise of AMD SEV-SNP

At first glance this looks like another cloud isolation story. Then you get to the part where a rogue host can read protected VM data, and in one case even alter memory and forge the attestation report.

A lot of cloud security news starts the same way: a boundary that was supposed to separate one layer from another did not hold up under pressure.

That sounds routine.

The part that is not routine is this: two newly disclosed attacks, Mior and Skyfall, let a malicious host read data inside AMD SEV-SNP virtual machines, which are designed specifically to keep the server operator out. And Mior reportedly goes further by modifying VM memory and faking the signed report that is supposed to prove what software the VM is running.

That cuts at the main selling point.

Why SEV-SNP exists in the first place

A virtual machine normally runs on hardware owned and controlled by someone else. In a cloud setup, that means the host system and its hypervisor sit beneath your VM and usually have enormous power over it.

AMD SEV-SNP is meant to change that relationship. The basic idea is to encrypt a VM's memory and add integrity protections so the host cannot casually peek inside or tamper with what is there.

Think of it like renting a storage unit where the facility owner still controls the building, the doors, and the cameras, but your unit is supposed to be locked in a way they cannot open. The whole point is reducing how much you must trust the landlord.

Attestation is part of that promise. It is the signed proof a VM can present to show what environment and software state it is running with, so another system can decide whether to trust it.

If that proof can be forged, the trust chain gets shaky fast.

What makes Mior and Skyfall serious

The public details so far say both attacks require control of the server's hypervisor. Mior also requires control of the server's BIOS.

That means these are not drive-by internet attacks. They are host-level attacks against the very layer confidential computing is supposed to defend against.

Skyfall is bad enough on its own if it allows a rogue host to read secrets from inside a protected VM. Secrets are often the whole reason teams use these features at all:

  • encryption keys
  • tokens and credentials
  • customer data in memory
  • model weights or proprietary code

Mior is the one that really changes the conversation. Reading memory is one class of failure. Being able to modify memory and fake the signed report is another.

At that point, it is not just about secrecy. It is also about whether the VM can still prove it is genuine.

That matters for systems that make high-stakes trust decisions based on attestation. A service may only hand over a key if the VM proves it booted the expected code. If an attacker can fake that report, they may be able to get the key anyway while running something else.

So the issue is larger than "can someone spy on RAM." It touches the mechanism people use to automate trust in cloud environments.

The practical read, without the hype

It would be easy to overstate this and imply every SEV-SNP deployment is instantly worthless. That is not the right read.

These attacks reportedly need deep control of the host. One also needs BIOS control. That is a high bar.

But confidential computing features like SEV-SNP exist for high-trust and high-value cases where the host operator is exactly the party you are trying to limit. So if the host can break the boundary anyway, even under specific conditions, that is a real architectural problem.

This is the awkward part of hardware-backed trust: most people hear "protected by hardware" and assume that means settled and simple. It is neither. The protection is only as strong as the chain beneath it, including firmware, boot process, and the logic used to prove the VM is in a good state.

When an attack can reach into both memory protection and attestation, it raises a harder question than a normal software bug does. Not just "is there a patch," but also "what assumptions were we making about who we had to trust?"

What cloud customers and defenders should do now

If you rely on SEV-SNP for sensitive workloads, now is the time to review where you trust attestation and what secrets are released based on it.

A few practical steps help:

  • Identify workloads that depend on SEV-SNP to protect data from the host itself.
  • Review any systems that automatically release keys, tokens, or data after attestation succeeds.
  • Treat attestation as one control, not the only control.
  • Limit the lifetime and scope of secrets kept in VM memory.
  • Ask your cloud or infrastructure provider how they protect hypervisors and firmware, including BIOS update and access controls.
  • Watch for vendor guidance on mitigations, firmware, BIOS, or platform updates tied to these attacks.
  • Keep separate monitoring around unusual key use or data access, because confidential computing does not replace detection.

If you are designing systems around confidential VMs, avoid building a setup where one successful attestation check grants long-lived access to everything. Short-lived credentials and layered trust still matter.

My read is simple: Mior and Skyfall are notable because they target the exact trust boundary SEV-SNP is supposed to create. If the host can read the VM, and in one case even change it and fake the proof, then anyone using that boundary for serious isolation needs to re-check their assumptions right now.