Researchers from MIT have unveiled a new speculative execution attack, dubbed TONTOU (Time-of-Neutralization to Time-of-Use), that circumvents existing defenses against Spectre v2. The attack, developed by Daniël Trujillo and Mengjia Yan of MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), exploits a brief window between when branch predictor states are neutralized and when they are subsequently used, allowing an attacker to re-poison the predictor.
The TONTOU attack targets mitigations designed to sanitize branch predictor states before sensitive code execution, a critical defense against Spectre-style vulnerabilities. Different chipmakers employ varying neutralization strategies: Intel's Enhanced Indirect Branch Restricted Speculation (eIBRS) clears predictors during context switches, while AMD's Safe RET, introduced after the 2023 Inception attack, focuses on sanitization immediately before a protected branch. Both approaches, however, share the assumption that attackers cannot alter branch predictor states within the "post-neutralization window."
Trujillo and Yan's research demonstrates that this assumption is flawed. They developed an "interrupt injection" primitive where an unprivileged program schedules high-frequency timer interrupts. The goal is for one of these interrupts to land precisely within the post-neutralization window. If successful, the interrupt handler executes after sanitization but before the victim branch is used, allowing it to re-poison predictor structures like the return stack buffer (RSB) or branch history buffer (BHB). This re-poisoning can then cause a protected branch to speculatively jump to a disclosure gadget, leaking kernel data via a side channel.
The researchers successfully demonstrated TONTOU on both Intel and AMD processors. Their tests included Intel Cascade Lake Refresh and Arrow Lake chips, as well as AMD Zen 2 and Zen 4 architectures. While the attack concept was proven across these platforms, a complete end-to-end exploit was developed specifically for AMD Zen 2, partly due to the more stringent software conditions required for the Intel attack.
The end-to-end exploit on AMD Zen 2 was used to bypass Linux's kernel address space layout randomization (KASLR), a crucial security feature. By precisely injecting interrupts to poison the RSB, the researchers consistently broke KASLR in all ten test runs. In five of these runs, they were also able to locate and leak the contents of `/etc/shadow`, which contains root password hashes. Each successful end-to-end attempt took approximately 18 minutes.
The researchers emphasize that their demonstration used a stock Linux kernel with all default mitigations enabled and no special system configurations or inserted modules. This indicates that any system where unprivileged code can schedule timers and shares the kernel with a victim could be vulnerable. Multi-tenant container platforms, for instance, would fall into this category, potentially allowing user-space programs to leak memory from the shared kernel.
Following their findings, the researchers engaged Intel, Arm, and AMD. AMD has committed to addressing the issue through kernel patches. Intel, however, stated it would not develop further mitigations, citing the numerous factors required for real-world exploits, such as the availability of disclosure gadgets, though it did award a bug bounty. Arm indicated that TONTOU's interrupt injections fall under "passive leakage," which it does not actively protect against.
The researchers hope their work will encourage further investigation into interrupt injections and TONTOU attacks, leading to more robust mitigations against Spectre-style exploits. Despite the complexity and time required for a successful attack, the demonstration proves the practicality of re-poisoning branch predictors during the post-neutralization window.
