Researchers Daniël Trujillo and colleagues at MIT CSAIL disclosed a new speculative-execution attack in early August 2026 that can bypass recent hardware mitigations against Spectre v2 on both Intel and AMD processors, according to reporting by BleepingComputer and The Hacker News. The team built a working exploit and used it to acquire a Linux system's password file, MIT's own news office reported.
The technique, nicknamed TONTOU, times a hardware interrupt to land in a narrow gap between the moment a processor sanitizes its branch predictor as a Spectre v2 defense and the moment the kernel actually uses that predictor, allowing an unprivileged program to re-poison the predictor after the protection has already run.
Why it matters
Spectre v2, first disclosed in 2018, has been the subject of repeated rounds of hardware and microcode mitigations from Intel and AMD, and defenders had largely treated the branch-predictor isolation techniques deployed since as settled. TONTOU shows those defenses can be bypassed not by breaking the sanitization step itself but by exploiting the timing window immediately after it — a reminder that mitigations addressing the wrong point in the execution pipeline can leave systems exposed even when every documented patch has been applied.
Because the attack only requires unprivileged local code execution — not physical access or existing kernel privileges — it is relevant to any multi-tenant environment where untrusted code runs alongside sensitive processes, including shared cloud infrastructure and any system that runs third-party or sandboxed code locally.
How it works
According to the summary published by security newsletter Suriq, TONTOU re-poisons the branch predictor after Spectre v2 defenses run to leak kernel memory, and it requires local code execution rather than a purely remote vector. The Hacker News' technical description frames the core mechanism as interrupt injection: an unprivileged Linux program deliberately times a hardware interrupt to land in the gap between the processor sanitizing the branch predictor and the kernel relying on that sanitized state, allowing the attacker to insert poisoned branch-prediction data into that gap. This report does not reproduce exploit code or step-by-step attack instructions.
Evidence
The disclosure is corroborated by MIT's own CSAIL press office, which credited the finding to researchers including Daniël Trujillo and described the demonstrated exploit's ability to acquire a Linux password file. BleepingComputer's independent reporting states the researchers found a way to bypass recent mitigations for Spectre v2 on processors from both AMD and Intel that rely on sanitizing or isolating branch-predictor state. Suriq's summary states AMD has patched the issue while Intel had not as of its August 7 report, though this report could not independently confirm Intel's current patch status beyond that account.
What defenders should do
Linux administrators, particularly those running multi-tenant or shared-hosting environments on Intel or AMD hardware, should monitor for microcode and kernel patches addressing TONTOU from their CPU vendor and Linux distribution, and apply them promptly once available. Organizations that cannot immediately patch should evaluate whether disabling simultaneous multithreading or applying kernel-level branch-predictor isolation flags — standard Spectre-class mitigations — offers interim protection, while recognizing those steps carry a performance cost and may not fully close this specific timing gap.
What happens next
Expect Intel and AMD to issue microcode or kernel-level guidance directly addressing the interrupt-injection technique, following the pattern of prior Spectre-class disclosures where hardware vendors iterate on mitigations over months rather than issuing a single fix. Given that TONTOU is at least the latest in a long line of Spectre v2 variants discovered since 2018, security researchers are likely to continue probing the same class of timing gaps in branch-predictor isolation for further bypasses.