AMD's Secure Cloud VMs Can Leak Your Data

New attacks let cloud hosts read and modify data inside supposedly secure AMD VMs, breaking a core privacy promise.

Two new attacks, Mior and Skyfall, have been disclosed that target AMD's Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP). According to reporting from The Hacker News, these vulnerabilities allow a malicious cloud host operator to bypass the technology's core protections. SEV-SNP is designed to create a trusted execution environment, or "confidential VM," where data and code are encrypted and isolated even from the hypervisor that manages the server. These attacks effectively break that promise, enabling a rogue host to read data from within a supposedly secure VM. This discovery challenges one of the foundational security models for confidential computing in modern cloud environments.

The two attacks, while achieving similar outcomes, operate differently. The Skyfall attack exploits a specific hardware vulnerability to leak data from the VM's memory. Mior is more complex and powerful, requiring the attacker to have control over both the server's hypervisor and its BIOS. With this level of access, an attacker can not only read a VM's memory but also modify it in real-time. Crucially, Mior can also forge the cryptographic attestation report. This report is a signed proof that is supposed to guarantee to a user that their VM is running the correct software on a secure, untampered system. By faking this report, the attack can deceive the VM's owner into believing their environment is secure while it is actively being compromised.

This is not the first time confidential computing technologies have faced scrutiny. Both AMD's SEV and Intel's SGX have had a history of vulnerabilities discovered by security researchers. These findings are part of a continuous cat-and-mouse game between hardware vendors building security primitives and researchers testing their limits. What makes these attacks significant is that SEV-SNP was AMD's enhanced version, designed to address weaknesses in earlier iterations of the technology. The discovery of Mior and Skyfall demonstrates that even these hardened solutions are not infallible. It reinforces the principle that security is a process, not a product, and highlights the immense difficulty of building a truly impenetrable barrier between a cloud guest and a malicious host with deep system access.

For organizations using or considering AMD SEV-SNP for sensitive workloads, this news is a critical reminder that "confidential computing" is not a silver bullet. While it provides a powerful layer of defense, it does not eliminate the need for traditional security best practices like defense-in-depth, strict access controls, and continuous monitoring. The attack's prerequisite of a compromised hypervisor or BIOS means the immediate threat is from a malicious or deeply compromised cloud provider, not a casual attacker. Users of AMD-based confidential VMs should monitor communications from their cloud providers and AMD for patches and mitigation guidance. This incident will likely spur further research into hardware-level security and push vendors to develop even more resilient isolation mechanisms for the next generation of cloud infrastructure.

These attacks undermine the fundamental promise of confidential computing: protecting data in use, even from the cloud provider. For teams relying on AMD SEV-SNP for sensitive workloads, this means their data isolation model may be flawed, requiring an urgent re-evaluation of their security posture and trust assumptions.

Companies using AMD-based confidential VMs to meet compliance or protect trade secrets now face a critical risk. A compromised cloud host could lead to data breaches, intellectual property theft, and regulatory fines, eroding customer trust in services built on this technology.