Improved Retbleed Exploit Leaks Host Memory from AMD Zen 2 Systems

Researchers demonstrated an improved Retbleed exploit against AMD Zen 2 processors that can read arbitrary host physical memory from an unprivileged, sandboxed process or from a KVM virtual machine with guest-kernel privileges. Published by Google Bug Hunters, the research showed leakage of approximately 13 KB/s with 100% accuracy under ideal, idle-machine conditions—roughly three times faster than the original exploit. The technique can expose other processes’ data and potentially memory belonging to other virtual machines; the report describes research demonstrations, not an observed malicious campaign.

Building on Spectre-era speculative-execution and cache side-channel research, the exploit combines branch-predictor manipulation, address-randomization bypasses, speculative return-oriented programming, and cache timing. It requires prior arbitrary native-code execution and knowledge of relevant kernel gadget and return offsets, rather than providing an initial compromise path. Linux enables software mitigations by default, although these impose performance costs; Google’s article states that AMD addressed Retbleed beginning with Zen 3. Organizations operating Zen 2 infrastructure, particularly systems running untrusted workloads, should verify that applicable kernel mitigations remain enabled and assess exposure across sandbox and virtualization boundaries.