Mior and Skyfall: New Attacks Break AMD SEV-SNP Confidential VM Protections
Confidential computing promises a simple but powerful guarantee: even the operator of the physical server cannot peek inside your virtual machine. Two newly disclosed attacks, dubbed Mior and Skyfall, challenge that guarantee for AMD's SEV-SNP (Secure Encrypted Virtualization–Secure Nested Paging) technology — and one of them goes further than reading secrets.
According to reporting from The Hacker News, both techniques allow a malicious host to extract data from inside SEV-SNP-protected VMs. Mior, the more severe of the pair, can additionally tamper with a VM's memory and forge the signed attestation report that is supposed to prove which software the guest is running.
What SEV-SNP Is Supposed to Do
AMD's SEV family encrypts VM memory so that the hypervisor and the underlying host cannot read it. SEV-SNP added memory integrity protections and a hardware-rooted attestation mechanism: a guest can request a signed report describing its launch state, then use that report to convince a remote party that it is running on genuine, untampered hardware with the expected firmware and configuration.
That attestation report is the linchpin of most confidential-computing trust models. Workloads handling regulated data, key material, or multi-tenant AI inference rely on it to decide whether a remote machine is safe to entrust with secrets. An attacker who can forge that report undermines not just confidentiality but the entire trust decision.
How the Attacks Differ
Both attacks require a high level of privilege on the host — specifically, control of the hypervisor. That is a meaningful caveat: these are not remote, unauthenticated exploits against a guest. They target the threat model SEV-SNP was explicitly designed to defend against — a curious or outright hostile cloud provider, or an attacker who has already compromised the host management layer.
Mior raises the bar in two ways. First, it can modify memory inside the protected VM rather than merely reading it. Second, it requires control of the server's BIOS in addition to the hypervisor, which narrows the pool of realistic attackers but also points to the firmware layer as part of the attack surface. With those capabilities, Mior can fabricate an attestation report that misrepresents the guest's software state — meaning a remote verifier could be convinced that a compromised environment is trustworthy.
Skyfall, by comparison, is described as a read-oriented attack. The public summary stops short of detailing its full mechanics, but its practical impact is the same category of failure: data that SEV-SNP promised to keep sealed away from the host becomes visible to the host.
Why This Matters for Cloud Customers
The uncomfortable implication is that confidential VMs should not be treated as an absolute boundary against infrastructure operators. Organizations that moved sensitive workloads into SEV-SNP environments on the assumption that the cloud provider was cryptographically excluded from the data path now need to revisit that assumption.
Several practical consequences follow:
- Attestation is only as strong as the firmware and hypervisor beneath it. If those layers are compromised, the signed report may no longer reflect reality. Verifiers should consider pinning expected measurements and treating unexpected changes as a hard failure rather than a warning.
- Defense in depth still matters inside the enclave. Application-level encryption, secrets management with short-lived credentials, and minimizing the blast radius of any single workload reduce exposure when a platform guarantee weakens.
- Provider trust is a risk input, not a checkbox. Customers should ask providers about patch status, firmware update cadence, and whether mitigations for these specific attacks are available in their regions and instance families.
The Broader Pattern
Mior and Skyfall fit a recognizable trend: as confidential computing moves from research curiosity to production feature, the attack research is maturing alongside it. Earlier generations of SEV flaws were largely about cryptographic or microarchitectural weaknesses in the encryption itself. These attacks target the surrounding trust machinery — attestation, firmware, and the hypervisor's control over guest state.
That shift suggests defenders should focus less on the cipher and more on the chain of trust that surrounds it. Firmware integrity, hypervisor hardening, and rigorous attestation verification are where the practical battles will be fought.
Recommended Actions
Security teams running SEV-SNP workloads should track AMD's guidance and their cloud provider's advisories for mitigation status, validate that attestation verification in their pipelines rejects anomalous reports, and review which workloads genuinely depend on host-exclusion guarantees versus which merely benefit from them. Where the guarantee is load-bearing, assume it can fail and layer controls accordingly.
Confidential computing remains a valuable tool. But Mior and Skyfall are a reminder that "the host can't see your data" is a claim with conditions attached — and those conditions deserve scrutiny.
Source: The Hacker News
