Researchers disclosed a new attack called GPUThor that can break key memory-safety protections on some NVIDIA GPUs and, in testing, helped an unprivileged program gain root access on the host system. The Rowhammer-style attack targets Ampere-class NVIDIA GPUs with GDDR6 memory, including RTX A4000, A4500, A5000, A6000, and likely A100-class server GPUs; it bypasses SECDED error-correcting code (ECC) and Target Row Refresh mitigations by using a non-uniform hammering pattern. NVIDIA published guidance on August 21 recommending SYS-ECC and IOMMU/DMA isolation.
Why it matters: Organizations using affected NVIDIA workstation or server GPUs for AI, cloud, or shared compute workloads could face data corruption, GPU resets, or host compromise from local code running on the system. This is urgent for defenders running multi-user GPU environments: review NVIDIA's advisory, enable the recommended mitigations, and limit untrusted CUDA workloads.
Bill Toulas
2026.08.26
100% relevant
This article appears to establish a new tracked event: a newly disclosed GPUThor attack against NVIDIA GPUs, with original research details and reference to NVIDIA's newly published mitigation advisory.
← Back to all stories