Security researchers from KU Leuven, ETH Zurich, Durham University and Google have demonstrated a hardware attack, dubbed DDRop, that can break the confidentiality and integrity guarantees of "confidential computing" systems by exploiting a design flaw in DDR5 memory encryption. Using a cheap custom circuit board wired between the processor and memory module, an attacker with physical access can silently drop write operations on the encrypted memory bus, tricking a protected virtual machine into computing on stale but validly decrypted data. This matters because it undermines trust in confidential computing offerings from major cloud providers, which rely on Intel TDX, Scalable SGX and AMD SEV-SNP to keep tenant data isolated even from the host operator.

The interposer costs under $200 to build, a fraction of the roughly $170,000 in lab equipment needed for earlier, weaker attacks, and the researchers say it works at full memory-bus speed, unlike prior passive eavesdropping methods that had to slow the bus down. In proof-of-concept tests against an Intel TDX server, the team forced a protected VM into debug mode to read its private memory in plaintext and even forged attestation reports so a compromised VM would appear trustworthy, both attacks succeeding reliably in under two minutes without crashing the machine. The researchers say there is no simple software or hardware fix, since scalable memory encryption inherently trades cryptographic freshness for the ability to protect large memory pools; Intel acknowledged the disclosure in a security bulletin published on Monday, and the team plans to release the interposer design as open-source hardware.