MIT researchers find a tiny CPU security gap attackers can exploit
The TONTOU attack uses precisely timed interrupts to slip past processor defences and recover protected data from a Linux system.
Modern processors are brilliant at guessing what software will do next. That shortcut is also becoming a rather awkward place to hide an attack. MIT researchers have demonstrated a new class of processor exploit that slips through a tiny gap in security defences and, in a working test, helped recover a Linux system’s password file.
The research, reported by MIT News and MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), is called TONTOU — short for Time-of-Neutralization to Time-of-Use. It targets the split second between a processor clearing its prediction machinery and actually using it.
The tiny gap inside a modern CPU
Processors use branch prediction to keep work moving. When a program reaches a fork in the road, the chip guesses which path it will take and starts early. A correct guess saves time. A wrong one is discarded, but the activity can leave traces behind — the basic territory made famous by the Spectre vulnerabilities disclosed in 2018.
Chipmakers have introduced defences that wipe or isolate this prediction state. MIT’s Daniël Trujillo and Mengjia Yan found that the cleanup and the moment the predictions are used cannot happen at precisely the same instant. There is always a small time-of-neutralization-to-time-of-use window, and code running inside it can dirty the prediction machinery again.
Interrupt injection makes the timing work
The researchers’ technique, called interrupt injection, uses ordinary processor interrupts — the little detours triggered by timers, network traffic and hardware. By tuning a timer with enough precision, an attacker can make an interrupt land during the vulnerable window and contaminate the state just before it is used.
The team tested four processor generations from Intel and AMD and triggered mispredictions on both. On Intel, the attack defeated two protections: one implemented in software for older processors and another built into the silicon of newer parts. One Intel generation resisted while another did not, suggesting that similarly named protections can behave rather differently in the real world.
AMD’s saferet defence leaves a window only two instructions wide, normally lasting tens of nanoseconds. The researchers slowed the processor at the crucial moment to make that tiny target easier to hit.
From a bad guess to /etc/shadow
This was not just a laboratory graph. On an AMD system running a current Linux kernel, the researchers first bypassed address-space randomisation, which normally makes protected memory harder to locate. They succeeded in all 10 attempts, taking about nine minutes each.
That let them read protected memory at roughly five bytes per second — glacial by ordinary computing standards, but enough to find and copy /etc/shadow, the file containing the system’s root password hash, in half of their attempts. It is a useful reminder that “slow” does not mean “harmless” when the target is a secret that should never leave protected memory.
The fix is not as simple as wiping harder
The paper proposes cleaning the prediction machinery again when an interrupt finishes. That appears workable on AMD. On Intel, the researchers warn it could make the attack more reliable because the exploit depends on the interrupt leaving a consistent state, not on leaving one particular state behind. Newer Intel processors also include a dedicated instruction that appears to help.
Blocking interrupts during the vulnerable window would close the door, but the performance cost would likely be too high for everyday systems. Security patches always have a bill attached; this one could be paid in processor time.
Trujillo and Yan notified AMD and Intel in February and contacted Linux kernel maintainers in March. AMD coordinated warnings to cloud providers and released a patch that mitigates the attack. The practical advice is refreshingly unglamorous: update your operating system, especially on machines handling sensitive data.
Our opinion
TONTOU is the sort of security problem that sounds microscopic until you remember what lives on the other side of the gap. The exploit is technically demanding, slow and not a magic “hack every computer” button, but the demonstrated access to a root password hash proves the risk is more than academic. AMD’s patch is the right immediate response; everyone else should treat processor and kernel updates as security maintenance, not optional housekeeping.