Researchers detail Interrupt Injection side-channel on AMD and Intel processors, exposing /etc/shadow contents despite eIBRS and Safe RET mitigations

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have unveiled findings on 6 August 2026, a new CPU side-channel technique that cuts straight through existing Spectre v2 defenses on both AMD and Intel chips, showing that it is possible to pull Linux password hashes out of kernel memory despite modern mitigations.

The work, described on their website, targets Spectre v2-style branch prediction protections that vendors have promoted as core safeguards against speculative execution attacks.

The attack primitive is named TONTOU (“Time-of-Neutralization to Time-of-Use”) and it focuses on a subtle timing gap in how processors handle branch predictor cleanup. Contemporary Spectre v2 countermeasures such as Intel’s Enhanced Indirect Branch Restricted Speculation (eIBRS) and AMD’s Safe RET mechanism, operate by flushing or compartmentalizing branch prediction state so it cannot be abused by speculative execution.

However, as PhD student Daniël Trujillo and associate professor Mengjia Yan discovered, there is an unavoidable window between the instant the branch predictor is sanitized and the point at which those sanitized predictions are actually consumed by the CPU. TONTOU weaponizes that tiny interval using what the researchers call “Interrupt Injection”.

Interrupt Injection explained

By carefully arranging for a hardware interrupt to arrive precisely inside that neutralized-to-use gap, an unprivileged user-space process can re-poison the branch predictor after the defense has already executed, but before the CPU relies on the supposedly safe prediction data.

In a conversation with security outlet BleepingComputer, Trujillo explained that this enables an attacker with no special privileges “to read arbitrary memory from the system, including sensitive data such as hashed passwords.” The result is that protection schemes previously thought to block speculative branch mispredictions can be sidestepped by extremely fine-grained timing control delivered purely from local code.

In 10 independent trials, the exploit succeeded in identifying and extracting the contents of the /etc/shadow file — which on Linux holds the system’s password hashes — in five cases. On average, each successful run took about 18 minutes to complete from start to finish, illustrating that the leak is slow but feasible in realistic time frames for a determined local adversary.

The researchers also confirmed that the same TONTOU concept applies to Intel processors that ship with Spectre v2 defenses like eIBRS. Nonetheless, they note that practical exploitation on Intel hardware is more complex because it depends on additional software conditions and configuration details, making the proof-of-concept attack harder to reproduce than on the tested AMD Zen 2 system. That nuance is important: both vendors’ CPUs are theoretically exposed, but the real-world barrier to exploitation differs per platform.

Official responses

AMD has responded to the disclosure by acknowledging that the problem “appears to be associated” with how the Linux kernel integrates the Safe RET mitigation rather than being purely a consequence of the microarchitecture itself. In other words, the interaction between the operating system’s implementation and the CPU’s protective feature is what opens the timing window that TONTOU can reach. MIT CSAIL notes that AMD has now issued a patch aimed at neutralizing the attack path, and that mitigation is being delivered to end users through standard operating system updates, including Linux distributions.

The group intend to publish more technical detail when they present at USENIX Security 2026 in October. Their research highlights a persistent reality in hardware and microarchitecture security: even carefully engineered Spectre v2 mitigations can still leave exploitable timing gaps that attackers with only unprivileged local code execution can surgically target. As speculative execution defenses grow more intricate, TONTOU serves as a reminder that defenders must consider not just what state is sanitized, but exactly when and how that sanitization intersects with real-world CPU behavior.