Spectre v2 defenses on affected AMD Linux systems are not a one-time cleanup. A timed hardware interrupt can land in the tiny gap after the branch predictor is sanitized and before the kernel uses it, so a local attacker can re-poison it and keep speculative leakage alive on a machine that looks patched. MIT CSAIL demonstrated the bypass on AMD Zen 2 under Linux 6.14 with default Spectre v2 mitigations enabled. The attack leaked arbitrary kernel memory at 5.47 bytes per second with 91.97% accuracy, and the researchers showed it was enough to recover sensitive data such as /etc/shadow. Linux has already merged a fix for the Safe-RET interrupt-injection gap. That changes the containment model for shared AMD Linux hosts. The issue is not whether Spectre v2 mitigation is present, but whether it survives an interrupt at the wrong moment.
Part of the PlainSec briefing for 2026-08-07