Linux Safe RET Flaw on AMD Zen CPUs Enables Interrupt Injection Data Leaks

Researchers from MIT CSAIL disclosed an interrupt injection attack that can bypass Linux’s Safe RET mitigation for Spectre-class attacks on AMD Zen 1 through Zen 4 processors, potentially allowing information disclosure from kernel memory. AMD acknowledged the issue in its Return Address Security Bulletin and said the weakness stems from Linux’s implementation of the mitigation for Speculative Return Stack Overflow (SRSO) rather than the processor architecture alone. The flaw lets an attacker with code execution on a target system precisely time an interrupt to disrupt branch predictor sanitization and weaken the intended protection.

Linux has already patched the issue in the latest kernel Git code by addressing the handling of interrupts around the Safe RET mitigation path. The researchers reported demonstrating arbitrary kernel memory leakage on AMD Zen 2 systems running Linux with default Spectre v2 defenses, achieving 5.47 bytes per second at 91.97% accuracy and reading /etc/shadow in five of ten attempts. The work was reported by Daniël Trujillo of MIT CSAIL, presented at Black Hat USA, and accompanied by a proposed fix that adds a second predictor neutralization step on interrupt return; researchers also said related misprediction behavior was observed on some Intel systems, though AMD is the vendor that formally assigned and documented the vulnerability.