Confidential computing—the ability to execute workloads on untrusted infrastructure whilst keeping data encrypted and inaccessible to the host—has become a cornerstone of trust models for offshore and cloud deployments. Intel's TDX and AMD's SEV-SNP are the dominant platforms for this isolation. A newly disclosed attack, DDRop, demonstrates a practical path around these protections by attacking the memory subsystem itself.
How DDRop Breaks Memory Isolation
At the core, DDRop exploits the DRAM (dynamic random-access memory) interface between the processor and the memory controller. Rather than attempting to decrypt data or reverse cryptographic operations, the attack silently drops write commands to memory. The processor believes the write has succeeded, but the old encrypted data remains in place. When the processor reads from that address, it retrieves stale data—often decrypted during intermediate computation—as if it were current.
This attack surface exists because memory controllers traditionally don't authenticate or verify write operations. A device with physical access to the memory bus, or firmware-level control of the memory controller, can selectively intercept and suppress writes. The confidential computing processor remains unaware that data has not been persisted correctly, breaking the fundamental assumption of state continuity.
Threat Model and Access Requirements
The attack is not a remote exploit. It requires an attacker with either physical access to the server hardware or the ability to install a small circuit on the memory bus. In practise, this means the threat is most realistic during manufacturing, during supply chain compromise, or in scenarios where a datacenter operator or insider has malicious intent.
For hosted environments, this narrows the risk considerably. An attacker cannot execute DDRop over the network. They cannot do it from within a guest VM running on the same physical hardware. The attack requires control of the underlying infrastructure—essentially, the ability to modify hardware before or during deployment.
Implications for Confidential Computing Deployments
The DDRop disclosure raises important questions about the threat models vendors have assumed for confidential computing. Both Intel and AMD have marketed TDX and SEV-SNP as protections against untrusted hypervisors and hostile cloud operators. DDRop shows that isolation at the processor level does not automatically guarantee protection against attacks on the memory subsystem itself.
For infrastructure operators, the practical impact depends on deployment context. If you are running confidential workloads on rented hardware in a multi-tenant cloud, you must now consider not just software-level access by other tenants, but also hardware-level attacks by the facility operator or supply chain actors. If you operate your own datacenter and control hardware from procurement through deployment, the risk is primarily internal threat or supply chain compromise.
Mitigation is currently limited at the software layer. Detection might be possible through timing anomalies or by validating state against cryptographic commitments, but these approaches add overhead and are not guaranteed. The more durable fix likely requires changes to DRAM interfaces and memory controllers—for instance, adding authentication or write verification mechanisms—which would be implemented in future silicon generations.
What This Means for Your Infrastructure Decisions
Confidential computing remains valuable for protecting workloads against certain threat actors (software vulnerabilities, guest escape, hypervisor compromise). DDRop does not invalidate those protections. However, it should inform threat modelling. If your security posture assumes that a confidential computing platform protects against all forms of host compromise, you now have evidence that assumption is incomplete.
For organisations deploying sensitive workloads on offshore or privacy-focused infrastructure, the lesson is straightforward: understand what confidential computing actually isolates, and against which adversaries. Combine it with other controls—audited supply chains, secure boot, hardware attestation, access controls—rather than treating it as a complete solution. As hardware attacks mature, layered defences remain the most practical approach to reducing attack surface in untrusted environments.
