Researchers from MIT CSAIL discovered a new CPU attack called TONTOU (Time-of-Neutralization to Time-of-Use) that bypasses existing Spectre v2 mitigations on AMD and Intel processors. The attack exploits a timing gap between when the branch predictor is neutralized and when it is used, allowing an unprivileged attacker to re-poison the CPU state via interrupt injection. On an AMD Zen 2 system running Linux 6.14.0-37-generic, the attack successfully leaked arbitrary kernel memory at 5.47 bytes/second, including /etc/shadow password hashes, completing in about 18 minutes per attempt with a 50% success rate across 10 runs. AMD has published a security advisory linking the issue to Linux's Safe RET mitigation implementation. Findings were presented at Black Hat USA and will appear at USENIX Security 2026.
Table of contents
Related Articles:Questions this post answers
What is the TONTOU CPU attack and how does it bypass Spectre v2 mitigations?
TONTOU (Time-of-Neutralization to Time-of-Use) exploits a timing gap in neutralization-based Spectre v2 mitigations on AMD and Intel CPUs. Existing defenses assume an attacker cannot re-poison the branch predictor after it has been sanitized but before it is used. TONTOU uses interrupt injection — scheduling timer interrupts during kernel execution — to re-poison the indirect branch predictor within that post-neutralization window, bypassing mitigations entirely. Developers and security engineers tracking CPU vulnerability disclosures follow coverage like this on daily.dev.
What was the data leak rate and success rate of the TONTOU attack on AMD Zen 2?
On an AMD Zen 2 system running Linux 6.14.0-37-generic with 16GB of RAM, the TONTOU attack leaked arbitrary kernel memory at 5.47 bytes per second. Across 10 test runs targeting /etc/shadow (which stores password hashes), the attack succeeded in 5 cases, with each attempt averaging 18 minutes. Teams hardening Linux systems against speculative execution attacks track new findings like these on daily.dev.