AMD's Secure VMs Just Got Breached: What Mior and Skyfall Mean for You

·
Listen to this article~4 min
AMD's Secure VMs Just Got Breached: What Mior and Skyfall Mean for You

Two new attacks, Mior and Skyfall, let a malicious cloud host read data inside AMD SEV-SNP-protected VMs. One can even forge the attestation report. Here's what it means for anyone trusting confidential computing.

Two new attacks went public this week, and they're a big deal if you trust cloud providers to keep your virtual machines private. Both Mior and Skyfall let a malicious host read data sitting inside AMD SEV-SNP-protected VMs. That's the whole point of SEV-SNP, by the way — keeping the server operator locked out of your stuff. Now that promise looks a lot shakier. ### What SEV-SNP Was Supposed to Do AMD's Secure Encrypted Virtualization with Secure Nested Paging is designed to encrypt a VM's memory so even the people running the physical server can't peek inside. Think of it like a safe deposit box at a bank. The bank owns the building, but they shouldn't be able to open your box. Mior and Skyfall suggest the bank might have found a way in anyway. ### Mior: Read, Write, and Fake the Receipt Mior is the nastier of the two. It doesn't just read data from a protected VM — it can also modify the VM's memory. And here's the part that should make you sit up straight: it can fake the signed report that confirms which software the VM is actually running. That report is called attestation, and it's how you verify a remote server is trustworthy before you send it secrets. If an attacker can forge it, the whole chain of trust falls apart. ### Skyfall: A Different Door, Same Problem Skyfall takes a separate approach but lands in a similar place — pulling data out of VMs that were supposed to be sealed off. Neither attack is a walk in the park for an attacker, though. There are real barriers: - Both require control of the server's hypervisor, which is the software layer managing the VMs. - Mior also needs control of the server's BIOS, adding another layer of access an attacker has to pull off first. - That means these aren't drive-by exploits. You need a compromised or outright malicious cloud host. So who should worry? Mostly people running sensitive workloads on shared or rented infrastructure. If you're on a cloud provider you don't fully control, and you're relying on SEV-SNP as your last line of defense, this is worth a hard look. ### The Bigger Lesson Here's the thing about hardware-level security features: they're only as strong as their weakest assumption. SEV-SNP assumed the host could be hostile but couldn't break the encryption. These attacks chip at that assumption. As one researcher put it, the gap between "the host can't read your memory" and "the host can't read your memory yet" is doing a lot of quiet work in most threat models. What should you do right now? A few practical steps: - Check whether your provider has published guidance or patches for these specific attacks. - Don't treat any single security layer as absolute — defense in depth still matters. - Revisit your threat model if you're trusting confidential computing to protect you from a malicious host. Confidential computing isn't dead. But this week's news is a good reminder that "confidential" comes with fine print, and it's worth reading it carefully.