News Date: 2026-08-06
A newly disclosed processor attack called TONTOU shows that some widely deployed Spectre v2 mitigations may not fully close the door on speculative execution leaks. Researchers Daniël Trujillo and Mengjia Yan demonstrated that unprivileged code can exploit a small timing gap between the neutralization of branch-prediction state and its subsequent use by the operating system.
A Race Inside the Processor
Spectre v2 attacks manipulate a processor's branch predictor so that it temporarily executes instructions along an attacker-influenced path. Although those speculative operations are discarded, traces can remain in microarchitectural components such as the cache, allowing carefully written code to infer protected information.
Current defenses include Intel's enhanced Indirect Branch Restricted Speculation and AMD's Safe RET approach. TONTOU targets what the researchers describe as a Time-of-Neutralization to Time-of-Use window. By scheduling timer interrupts during kernel activity, malicious code can attempt to repopulate predictor state after it has been cleared but before a protected branch uses it.
In testing on an AMD Zen 2 system running Linux, the researchers extracted arbitrary kernel memory at approximately 5.47 bytes per second with nearly 92 percent accuracy. Their demonstration included recovering content from the system's shadow password file. The attack was also found to be possible on Intel processors, although additional software conditions made exploitation more complicated.
What IT Teams Should Do
- Follow kernel, processor vendor and Linux distribution advisories for updated mitigations.
- Limit the execution of untrusted local code, particularly on shared servers and development infrastructure.
- Review isolation controls for hosting environments where mutually untrusted workloads share physical processors.
- Keep firmware, microcode and operating system packages current.
The attack currently requires local code execution and significant technical precision, so it should not be treated like a remotely exploitable server vulnerability. However, I believe it is especially relevant to cloud providers, research clusters and other multi-tenant environments. TONTOU demonstrates that clearing sensitive processor state is not enough if an attacker can influence what happens in the moments immediately afterward. In my view, speculative execution security remains an architectural challenge rather than a problem that can be permanently solved by one generation of software mitigations.