Researchers Break AMD SEV-SNP Security to Read Encrypted Virtual Machine Memory Using a GPU

Blog WriterCybersecurity News - Original News Source is cybersecuritynews.com


An attack called Mior uses a graphics card to bypass AMD SEV-SNP, allowing access to confidential virtual machine memory, modification of guest memory, and making altered VMs appear to run trusted software. Tracked as CVE-2026-0462, the flaw carries a CVSS 4.0 score of 5.9.

AMD Secure Encrypted Virtualization with Secure Nested Paging, or SEV-SNP, is designed to protect data while a virtual machine is running. It encrypts each guest’s memory and uses hardware checks to prevent the hypervisor from accessing it.

This software runs virtual machines and prevents the hypervisor from reading or changing protected pages. That separation matters in the cloud because the provider controls the host, while the customer may be processing data the provider should never see.

The GPU did not crack the encryption or recover the key. Instead, the researchers showed that a malicious host could interfere with setting up the Reverse Map Table, which records who owns each memory page and who can access it.

Misconfiguration Hidden from PSP Memory Map (source : xca-attacks.github )
Misconfiguration Hidden from PSP Memory Map (source: xca-attacks.github )

In the lab, some of the AMD Platform Security Processor’s setup writes reached graphics card memory instead of system memory. That left holes in the table’s protection, undermining the checks that keep the host away from guest data.

AMD SEV-SNP GPU Attack

The ACM research paper shows that the Mior attack gained read and write access to a production-style guest VM with debugging disabled.

The attack also allowed a modified VM to spoof a legitimate guest’s launch measurement, potentially undermining attestation checks used to verify VM trustworthiness before handling sensitive data.

The team demonstrated Mior on a Zen 5 AMD EPYC 9135 server. The setup required control of the hypervisor and system firmware, along with access to a PCIe device such as a graphics card.

Mior attack overview (source : xca-attacks.github )
Mior attack overview (source: xca-attacks.github )

This was not a remote break-in by an ordinary cloud customer. It tested what a highly privileged host could do to a guest that SEV-SNP was built to protect.

Mior follows earlier findings covered by Cyber Security News. RMPocalypse also exposed a weakness in how the Reverse Map Table is set up.

At the same time, BadRAM exploited how AMD systems interpreted information from memory modules. Mior shows that the route a security processor’s writes take can matter just as much as the checks it performs.

In its November 10 security bulletin, AMD lists SEV-SNP as affected, but not the older SEV or SEV-ES modes. It lists SEV firmware fixes dated August 12, 2026, for EPYC 7003, 8004, 8005, 9004 and 9005 server processors.

PSP Mistakes GPU Memory for System Memory (source : xca-attacks.github )
PSP Mistakes GPU Memory for System Memory (source: xca-attacks.github )

AMD targets December 2026 for the related platform firmware releases to server makers. Several embedded EPYC models have separate target dates in 2027.

Server operators should check with their hardware vendor for the BIOS update that matches their model. Cloud customers should ask whether the fix is active on the hosts that run their workloads and check attestation reports against the values AMD lists for the updated firmware.

A guest operating system update alone cannot address a flaw in host firmware. The key question is not whether a patch exists, but whether the host running the workload has received it.

Give your SOC the alert context on the latest threats to stop incidents quickly -> Integrate Threat Intel from 16K SOCs

The post Researchers Break AMD SEV-SNP Security to Read Encrypted Virtual Machine Memory Using a GPU appeared first on Cyber Security News.