University of Toronto Reveals Rowhammer Attacks on Nvidia GPUs

Security researchers at the University of Toronto have demonstrated new Rowhammer attacks targeting Nvidia graphics processing units, enabling attackers to corrupt GPU memory and escalate privileges to achieve full control over affected systems. The techniques, detailed in studies set for presentation at the 47th IEEE Symposium on Security and Privacy in April 2026, affect high-end Nvidia hardware from the Ampere generation, including models like the RTX 3060, RTX 6000, and A6000, which use GDDR6 memory.

Rowhammer, a hardware vulnerability identified over a decade ago in dynamic random-access memory, causes bit flips in adjacent memory cells through repeated access, or "hammering," of targeted rows. While previously studied mainly on central processing units, recent work shows GPUs face similar risks. The new attacks build on 2025 research that first proved Rowhammer's practicality on GPUs, where it could degrade machine learning model accuracy by up to 80 percent.

In the primary attack, named GPUBreach, researchers induced bit flips to corrupt GPU page tables, structures that manage memory access. This allows an unprivileged process running a CUDA kernel, Nvidia's programming...