MIT Computer Science and Artificial Intelligence Laboratory researchers demonstrated a new attack that bypasses critical security defenses in processors from both Intel and AMD. The team exploited a fleeting gap between when a chip’s prediction machinery is cleared and when it’s used, successfully acquiring a Linux system’s password file. Researchers call this class of attack “TONTOU,” highlighting the exploitation of this timing vulnerability. “There is always a gap,” explains PhD student Daniël Trujillo and MIT Assistant Professor Mengjia Yan, revealing how interrupts are used to reliably inject code and contaminate the processor’s predictions.
TONTOU Attack Exploits Processor Prediction Gaps
Modern processors rely on prediction to accelerate tasks, yet a newly discovered vulnerability, dubbed “TONTOU: On the Exploitability of Time-of-Neutralization to Time-of-Use Windows,” exposes a critical timing gap in these defenses. The attack hinges on a fleeting window between when a chip’s prediction mechanisms are cleared and when those predictions are actively used; malicious code injected during this interval can compromise system security. Interrupts, routine tasks triggered by timers, network traffic, and hardware, provide the means for reliably injecting code into this vulnerable gap.
Ordinary programs can configure these timers, and the MIT team tuned them with precision to trigger a detour at the precise moment the processor attempts to use its predictions, contaminating the process. This technique proved effective against four processor generations from both Intel and AMD, revealing a systemic weakness in current security approaches.
On AMD systems, the researchers bypassed the saferet defense, which aims to clean prediction machinery immediately before use, by subtly slowing the processor to widen the vulnerable window. The team’s successful acquisition of the /etc/shadow file, containing the system’s root password hash, highlights the severity of the vulnerability. They achieved this on an AMD system running a current Linux kernel in approximately nine minutes, reading protected memory at roughly five bytes per second.
While a secondary cleaning of the prediction machinery after interrupt completion appears workable for AMD, it may inadvertently increase the attack’s reliability on Intel chips. The researchers have made their code publicly available and coordinated with AMD and Intel to release a patch mitigating the attack, now available through operating system updates.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
