New Processor Vulnerability Exploits Prediction Gaps, MIT Research Shows

A team of researchers at the Massachusetts Institute of Technology (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL) has uncovered a novel attack method against modern processors that exploits a fundamental timing gap in their prediction defenses. Led by PhD student Daniël Trujillo and assistant professor Mengjia Yan, the findings were presented at Black Hat USA and USENIX Security conferences, revealing vulnerabilities in Intel and AMD CPUs despite years of prior protective measures.

What Happened

Modern processors speed up computations by predicting program branches ahead of time, executing instructions speculatively before confirming which path is correct. When these guesses are wrong, the processor discards the work, but data traces linger, vulnerable to side-channel attacks like Spectre disclosed in 2018. To combat this, chipmakers implemented defenses that clean or isolate the processor’s branch prediction components.

The MIT researchers discovered that these defenses cannot simultaneously clean the prediction machinery and use it, creating a brief “gap” — sometimes only a few instructions long — during which the prediction unit can be contaminated again. Their new attack, dubbed “TONTOU,” leverages this gap by precisely triggering interrupts at vulnerable moments, a technique they call “interrupt injection.”

Testing multiple generations of Intel and AMD processors, the team consistently induced mispredictions. They demonstrated the attack’s capability by constructing a working exploit on an AMD system running a recent Linux kernel, managing to extract protected system data including parts of the “/etc/shadow” password hash file within minutes.

Key Facts

The research was conducted by MIT’s CSAIL under the guidance of Daniël Trujillo and Mengjia Yan and funded partly by the U.S. Air Force Office of Scientific Research and DARPA’s JUMP 2.0 program. The study analyzed four generations of Intel and AMD processors, focusing on the precise timing of interrupt handling and prediction machinery cleaning. The attack affected Intel chips despite their integrated defenses and broke AMD’s “saferet” mitigation, exploiting a vulnerability window as narrow as two instructions typically executed in tens of nanoseconds. The work was presented in 2024 at both Black Hat USA and USENIX Security conferences.

What This Means

This discovery reveals that even state-of-the-art processor defenses against speculative execution attacks are not foolproof, as they rely on an assumption about atomic cleaning and usage that does not hold in practice. For consumers and enterprises, especially cloud service providers and data centers relying on hardware-level security, this poses a risk where sensitive data could be leaked via side channels once more.

The practical extraction speed is slow, around five bytes per second, but enough to breach critical files such as password hashes, highlighting a fresh threat vector for attackers targeting system memory. The findings underscore the complexity of securing modern processors, where microarchitectural optimizations can inadvertently open subtle but exploitable vulnerabilities.

This challenges chipmakers to rethink or enhance interrupt handling and prediction clean-up mechanisms and calls for OS and firmware patches to mitigate the vulnerability, as well as a reassessment of the performance-security balance in speculative execution defenses.

What Remains Unclear

The researchers note that the ideal mitigation for Intel processors is less straightforward than AMD’s, as cleaning prediction machinery post-interrupt might paradoxically increase exploit reliability due to differing implementations of prediction state management. The full impact across other processor architectures besides Intel and AMD was not detailed and remains to be explored.

What Comes Next

Following disclosure, MIT coordinated with AMD and Linux kernel maintainers. AMD has since released a software patch addressing the issue, available through operating system updates. Future research will likely focus on exploring similar vulnerabilities in other chip designs and refining defensive architectures to close prediction windows completely without prohibitive performance costs.

Read more Cybersecurity stories on Goka World News.