TONTOU CPU Attack Tests Spectre Defenses On Linux Systems
Researchers showed a Time-of-Neutralization to Time-of-Use technique that can repollute branch prediction state after Spectre v2 mitigations and leak Linux kernel data in lab tests.
A new CPU side-channel attack can bypass recent Spectre v2 mitigations on Linux systems and leak privileged data, including password hashes, after an attacker gains the ability to run unprivileged code on a target machine.
BleepingComputer reported that Daniël Trujillo, a PhD student, and associate professor Mengjia Yan of MIT CSAIL found a way to exploit the short interval between the moment a branch predictor is neutralized and the moment a protected victim branch uses it.
The researchers call that interval Time-of-Neutralization to Time-of-Use, or TONTOU.
Spectre v2, also known as Branch Target Injection, abuses a processor's indirect branch predictor so the CPU speculatively executes instructions along an attacker-influenced path.
Intel and AMD mitigations such as Intel eIBRS and AMD Safe RET are designed to sanitize or isolate branch-predictor state before sensitive control flow executes.
TONTOU targets the assumption that the cleaned state cannot be usefully repolluted before the victim branch runs.
The researchers introduced a primitive that lets an attacker poison CPU state after neutralization but before use.
"An attacker without any special access to read arbitrary memory from the system, including sensitive data such as hashed passwords," Trujillo told BleepingComputer.
The attack uses interrupt injection.
Unprivileged user programs can schedule timer interrupts during kernel execution, causing the kernel to enter an interrupt handler.
That handler can then be used to poison microarchitectural state inside the post-neutralization window.
The researchers found that interrupts during that window can poison the processor's indirect branch predictor and enable attacks against all types of indirect branches.
Exploitation still requires several difficult steps: redirecting kernel control flow, aligning interrupts precisely with the post-neutralization window, and poisoning the predictor entry tied to the target indirect branch.
In tests on an AMD Zen 2 system running Linux 6.14.0-37-generic with 16GB of RAM, Trujillo and Yan demonstrated arbitrary kernel memory leakage at 5.47 bytes per second with 91.97% accuracy.
The leaked data included contents of /etc/shadow, the Linux file that stores password hashes.
Across 10 runs, the attack located and extracted the file in five cases.
Each attempt took an average of 18 minutes.
The attack was also tested on Intel processors, though the researchers found that additional software requirements made exploitation more complex.
On AMD systems, they combined interrupt injection with Inception, a previously disclosed attack that Trujillo helped develop, because passive Return Stack Buffer pollution was less reliable.
AMD published an advisory saying the interrupt-injection issue "appears to be associated" with how Linux implements the Safe RET mitigation against possible information disclosure attacks.
Trujillo and Yan presented the findings at Black Hat USA.
They are also scheduled to share details at USENIX Security 2026, which runs from October 27 to October 29.
