New Attack Bypasses Spectre v2 Defenses on Intel, AMD CPUs

Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- MIT CSAIL researchers Daniël Trujillo and Mengjia Yan disclosed INTERRUPT INJECTION — an unprivileged attack that times a hardware interrupt to land between a CPU's branch-predictor sanitization and its use, re-poisoning the predictor after the defense has run.
- On AMD Zen 2 running Linux 6.14 with every default Spectre v2 mitigation enabled, the exploit leaked arbitrary kernel memory at 5.47 bytes per second with 91.97% accuracy, successfully reading /etc/shadow in five of ten attempts.
- AMD engineers Borislav Petkov and David Kaplan landed the Linux kernel fix on June 2 under the commit "x86/bugs: Make Safe-RET robust against interrupt injection," one of two remediation routes the paper proposed.
- AMD's August 6 bulletin AMD-SB-7061 names Zen 1 through Zen 4 as affected but lists no CVE, no patch reference, and no kernel version, and neither the bulletin nor MIT's announcement points to the kernel commit.
- Intel paid a discretionary bug bounty but concluded mitigation is unnecessary; its existing guidance INTEL-SA-00598, last updated May 2025, makes no mention of interrupts anywhere.
- The attack class TONTOU (Time-of-Neutralization to Time-of-Use) exploits a two-instruction, six-byte window on Zen 2; interrupts landed in that window 5% to 12% of the time, with attacker-controlled registers around 2%.
- Mispredictions appeared on three of four tested machines: Zen 2 (0.75%), Intel Arrow Lake (0.22%), and Cascade Lake Refresh (0.037%); Zen 4 produced none, and no end-to-end Intel leak was demonstrated but researchers call it feasible.
Why it matters: AMD shipped a Linux kernel patch but published no CVE or patch reference, forcing admins to track the fix by commit name. Intel paid the bug bounty yet called mitigation unnecessary, despite the researchers' argument that an end-to-end Intel exploit is feasible by combining their primitive with prior disclosure-gadget work.
Ask SkimNews



