Researchers expose two host‑side attacks that break AMD SEV‑SNP VM confidentiality
Security researchers have disclosed two new techniques, dubbed Mior and Skyfall, that enable a compromised cloud host to extract data from virtual machines protected by AMD’s SEV‑SNP technology.
The attacks target the hardware‑based isolation that SEV‑SNP provides, which is intended to keep the guest operating system’s memory invisible to the underlying hypervisor and the cloud provider. By exploiting subtle flaws in the way the processor validates encrypted memory pages, the researchers demonstrate that a rogue host can decrypt and read the contents of a protected VM.
Mior, the first of the two methods, goes a step further by allowing the attacker not only to read but also to tamper with the VM’s memory. It can inject crafted data and manipulate the cryptographic attestation report that the VM generates, potentially disguising the presence of malicious modifications from the guest’s perspective.
Skyfall, the companion technique, focuses on extracting secrets without altering the guest state. It leverages a timing side‑channel combined with a crafted page‑table walk to coax the encrypted VM into revealing keys and other sensitive information to the host.
Both attacks rely on the fact that SEV‑SNP, while stronger than its predecessor SEV‑ES, still depends on a set of hardware‑enforced invariants that can be subverted when the host has full control over the page tables and can observe execution patterns. The researchers stress that the vulnerabilities do not stem from a software bug in the guest OS but from the design of the isolation mechanism itself.
The findings raise concerns for enterprises that rely on AMD‑based confidential computing to protect workloads in public clouds. Cloud providers may need to roll out firmware updates or redesign certain aspects of the SEV‑SNP attestation flow to mitigate the risk. Meanwhile, customers are advised to monitor vendor advisories and consider additional layers of encryption for the most sensitive data.
Comments (0)
Be the first to comment.
Join the discussion