
Mior and Skyfall Attacks Let Rogue Hypervisors Read AMD SEV-SNP VM Secrets
Researchers disclosed Mior and Skyfall attacks that let a malicious hypervisor read secrets from AMD SEV-SNP VMs. Mior also modifies memory and forges attestation, while Skyfall leaks saved state and a test disk-encryption key through transient scheduler attacks. The findings show confidential computing's host-isolation assumptions can fail during boot and startup failures.
Two attacks disclosed this week target AMD SEV-SNP, the confidential-computing feature meant to keep a server's operator out of a virtual machine. Mior, from ETH Zurich, lets a malicious hypervisor with BIOS control redirect startup writes from the chip's Platform Security Processor to a graphics card, leaving part of the Reverse Map Table, which records memory ownership, uninitialized. That lets the host rewrite protections, enable debugging and read or write VM memory. The team reported success in 1,000 attempts. Skyfall, from ETH Zurich and Google, leaks saved VM state through transient scheduler attacks.
The evidence is lab-based, not public exploitation. Mior was tested on one AMD EPYC 9135, a Zen 5 chip; researchers say Zen 3 and Zen 4 may be affected but could not test them because changing BIOS was not possible. Both Mior and Skyfall worked even with ciphertext hiding enabled, an optional SEV-SNP setting that normally blocks the hypervisor from reading encrypted VM memory. AMD has listed SEV firmware for Mior and microcode for Skyfall dated August 12, with platform firmware planned for November and December and fixes for embedded chips later.
The pattern is familiar in confidential computing: the strongest assumption breaks at boot. Mior corrupts the Reverse Map Table before it fully protects memory, while Skyfall abuses a failed VM launch to leave saved state cached for instruction-timing leaks known as transient scheduler attacks. Mior also copies a harmless VM's measurement into a malicious VM, spoofing signed attestation reports that enterprises use to verify what is running. The team says this sits inside AMD's stated threat model because AMD distrusts UEFI. That distinction matters: it is not a simple remote web exploit, but a host-level compromise of trusted infrastructure.
The stakes are highest for enterprises that treat SEV-SNP as a technical control against a compromised cloud host, especially for regulated data, disk-encryption keys and attestation-based zero-trust workflows where software checks are supposed to prove integrity. Mior can read and modify VM memory and fake software identity; Skyfall can recover saved registers and a disk-encryption key from a test VM. Ordinary VMs already trust the host, so the new exposure is the confidential VM promise. The fix path is firmware and microcode, not just cloud configuration; customers must get BIOS updates from server makers.
Enterprises relying on AMD SEV-SNP to hide secrets from cloud hosts are exposed because Mior can read, modify and forge attestation inside confidential VMs, while Skyfall can leak saved state and a disk-encryption key from a test VM.
Sources (1)
- [1]Original source(https://thehackernews.com/2026/11/new-mior-and-skyfall-attacks-let-rogue.html)