This post is also available in:
עברית (Hebrew)
Modern processors routinely predict what software will do next so they can begin executing instructions before they are actually needed. This technique, known as speculative execution, improves performance but has also created a persistent security problem: if attackers manipulate those predictions, they may be able to infer information from memory they should never be able to access.
Chipmakers and operating-system developers have introduced defenses that clear or isolate this prediction machinery before sensitive operations. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), however, have demonstrated a new attack showing that even extremely small timing gaps in those defenses can be exploited.
The attack class, called TONTOU, targets the brief interval between clearing the processor’s prediction state and actually using it. Even when that window lasts only a few processor instructions, the researchers found a way to insert activity that contaminates the prediction mechanism again.
Their technique, called INTERRUPT INJECTION, exploits a routine part of computer operation. According to TechXplore, processors constantly interrupt running programs to handle events generated by timers, network traffic and hardware. By precisely configuring a timer, the researchers could trigger an interrupt during the vulnerable window. Code executed while handling that interrupt modifies the processor’s prediction state, creating conditions that can lead to a misprediction when the original task resumes.
The researchers tested the technique on four generations of Intel and AMD processors and successfully generated mispredictions on both manufacturers’ hardware. On Intel systems, it bypassed software-based protection on older processors as well as hardware-based protection on one newer generation.
The company’s saferet defense presented an even narrower target. It clears prediction state immediately before use, leaving a window only two instructions wide, typically lasting tens of nanoseconds. The researchers nevertheless managed to target it by deliberately slowing processor execution at the critical moment.
To demonstrate the security implications, they developed a working exploit against an AMD computer running a current Linux kernel. The attack first defeated a protection that randomizes the operating system’s location in memory. It then leaked protected information at approximately five bytes per second. In half of their attempts, the researchers were able to locate and copy /etc/shadow, the file containing the system’s root password hash.
The findings are particularly relevant to defense, government and critical-infrastructure networks, where sensitive workloads increasingly share complex computing environments. Processor-level attacks can undermine security boundaries below conventional applications, meaning traditional antivirus or network monitoring may not address the underlying weakness.
The researchers disclosed the vulnerability to the companies’ maintainers. AMD has since released a mitigation that users can obtain through operating-system updates. The study highlights an uncomfortable reality of modern processor security: even defenses operating within nanoseconds can leave enough time for an attacker to find a way through.
