New GPUThor Rowhammer Defeats ECC on NVIDIA RTX A6000 to Gain Host Root Access
- ID
- 18523
- Status
- summarized
- Published
- 27 Aug 2026, 4:13 PM
- Fetched
- 27 Aug 2026, 6:32 PM
- Provider
- The Hacker News
- Category
- security
- Original URL
- https://thehackernews.com/2026/08/gputhor-rowhammer-defeats-ecc-on-nvidia.html
- Source URL
- https://feeds.feedburner.com/TheHackersNews
Summary
- Score
- 6.5
- Created
- 27 Aug 2026, 6:34 PM
- Tags
- Audience
- developersai_ml_learnerssaas_founders
What happened
University of Toronto researchers disclosed GPUThor, a Rowhammer attack that defeats System-Level ECC on NVIDIA Ampere workstation GPUs (RTX A6000, A5000, A4500, A4000), enabling privilege escalation to a root shell. The attack uses non-uniform hammering—activating the aggressor row far more than decoy rows—to bypass Target Row Refresh, which the researchers found likely fires only once every 72 refresh intervals rather than per interval. This contradicts NVIDIA's July 2025 security notice claiming System-Level ECC fully mitigates GPU Rowhammer.
Why it matters
If you operate multi-tenant GPU infrastructure or accept untrusted CUDA kernels (e.g., a GPU cloud, notebook-hosting SaaS, or shared inference platform), you should stop cross-tenant GPU sharing on these Ampere cards and monitor ECC error counters for anomalous bit-flip rates, since ECC no longer fully neutralizes the attack. Single-tenant shops running only their own trusted code are lower risk but should still restrict untrusted CUDA workloads.
Discussion angle
For founders running GPU-as-a-service or shared notebook platforms in Southeast Asia: does this change your isolation strategy, or is the 24-hour hammer window and unprivileged-CUDA prerequisite low enough risk to accept operationally?