Researchers from the Massachusetts Institute of Technology (MIT) have developed a novel speculative execution attack, dubbed TONTOU (Time-of-Neutralization to Time-of-Use). The attack demonstrates that by using precisely timed system interrupts, malicious code can circumvent existing protections designed to defend against Spectre v2 on both Intel and AMD processors.
The Vulnerability in Speculative Execution Defenses
Daniël Trujillo and Mengjia Yan, affiliated with MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), detailed their findings in a paper presented at Black Hat. Their research focuses on weaknesses in the mitigations designed to neutralize potentially hostile branch predictor states before sensitive code executes, a crucial defense against Spectre-style vulnerabilities.
Current industry mitigations vary slightly across chip manufacturers. For instance, Intel employs eIBRS to sanitize branch predictors during a context switch, while AMD utilizes Safe RET—a technique Trujillo co-authored in 2023—which focuses on cleaning the predictor state immediately preceding a protected branch execution. The researchers refer to these mechanisms as entry neutralization and in-place neutralization, respectively.
The core defense assumption these mitigations rely on is that the “post-neutralization window”—the time between when the predictor state is sanitized and when the victim branch is used—is entirely secure. The new TONTOU attack proves this assumption false, showing how attackers can actively re-poison the branch predictor within this narrow window.
Attack Mechanism and Scope
To execute the attack, the researchers developed a technique called “interrupt injection.” This method involves an unprivileged program scheduling high-frequency timer interrupts, aiming for one to fire during the vulnerable post-neutralization window. By triggering an interrupt at this precise moment, the attacker can redirect the system’s control flow. This allows a malicious interrupt handler to execute after the initial sanitization process but before the protected branch is utilized.
The interrupt handler then has the ability to re-poison key predictor structures, such as the return stack buffer (RSB) or the branch history buffer (BHB). This re-poisoning causes the protected branch to speculatively jump to a disclosure gadget, enabling the leakage of sensitive kernel data through a side channel.
The researchers confirmed that their tests demonstrated the viability of the TONTOU attack across both Intel and AMD-based Linux environments. They tested the exploit on AMD Zen 2 and Zen 4 chips, as well as Intel Cascade Lake Refresh and Arrow Lake processors. While building a complete end-to-end exploit was easiest on Zen 2, the demonstration showed that the attack works even when running on a stock Linux kernel, without custom modules, and with all default mitigations enabled.
In a practical demonstration, Trujillo and Yan successfully identified the exact injection point needed to corrupt the RSB. This led to a successful breach of Linux’s Kernel Address Space Layout Randomization (KASLR). Across ten total runs, the researchers consistently broke KASLR, and were able to successfully locate and leak the contents of the critical system file, `/etc/shadow` (which contains the root password hash), in five of those attempts.
This vulnerability makes multi-tenant container platforms, where unprivileged user-space programs share a kernel with a victim, particularly susceptible to the attack.
Industry Response and Implications
The research has prompted responses from major hardware manufacturers. After presenting their findings, the researchers engaged representatives from Intel, Arm, and AMD.
- AMD: Committed to addressing the issue through future kernel patches.
- Intel: Declined to implement new mitigations, citing the overwhelming number of real-world variables, though the company awarded a prize through its bug bounty program.
- Arm: Stated that TONTOU’s interrupt injections fall under “passive leakage,” which the architecture does not “actively protect against.”
The authors concluded that while speculative side-channel attacks remain highly complex to execute, their work highlights the need for further investigation into interrupt injections and the development of more robust defenses against Spectre-style exploits.
