Ethical Hacking News
A new Rowhammer attack, dubbed "GPUThor," has been discovered, which reveals the limitations of NVIDIA's error correction mechanism in their high-end GPUs. The attack, which was made possible by non-uniform hammering of DRAM banks, has been found to induce bit flips on each DRAM bank, thereby corrupting the data stored in the GPU's memory. The attack has significant implications for the security and stability of NVIDIA's high-end GPUs and highlights the need for stronger multi-bit error correction and in-DRAM defenses.
The "GPUThor" Rowhammer attack exploits vulnerabilities in NVIDIA's error correction mechanism (ECC) in their high-end GPUs.The attack induces bit flips on each DRAM bank, corrupting data stored in the GPU's memory.The attack was tested on various NVIDIA GPUs, including the RTX A6000, RTX A5000, and found to be successful on all.The attack allows privilege escalation and exploits single-error-correct, double-error-detect (SECDED) ECC, resulting in silent data corruption (SDC).NVIDIA has released a security notice with guidance on how to mitigate the attack.The attack highlights the need for stronger multi-bit error correction and in-DRAM defenses.
The world of cybersecurity has recently witnessed the emergence of a novel Rowhammer attack, dubbed as "GPUThor," which has exposed the vulnerabilities of NVIDIA's error correction mechanism (ECC) in their high-end GPUs. The attack, which was disclosed by researchers at the University of Toronto, has been found to defeat ECC, a mitigation mechanism that was previously recommended by NVIDIA to prevent GPU Rowhammer attacks.
The discovery of the GPUThor attack has significant implications for the security and stability of NVIDIA's high-end GPUs, which are widely used in various industries such as gaming, professional graphics, and scientific simulations. The attack, which was made possible by the non-uniform hammering of DRAM banks, has been found to induce bit flips on each DRAM bank, thereby corrupting the data stored in the GPU's memory.
The researchers behind the GPUThor attack used a combination of techniques to induce bit flips on each DRAM bank. They first identified that repeated accesses issued inside a single warp were merged at the memory controller into a single DRAM activation. This allowed them to distribute the hammering kernels accordingly and exploit the limited Target Row Refresh (TRR) defense. The researchers also found that TRR on these GDDR6 parts likely applied about once every 72 refresh intervals rather than once per interval, which enabled them to build a six-interval pattern around that schedule.
The attack was tested on various NVIDIA GPUs, including the RTX A6000, RTX A5000, RTX A4500, and RTX A4000, which were found to be vulnerable to the attack. The researchers found that the RTX A5000 was the most susceptible at 377,552 flips per gigabyte, which is 23,597 times the 16 flips per gigabyte reported for GPUHammer's attack. The researchers also found that the attack produced 72,000 to 377,000 bit flips per gigabyte with ECC disabled.
The attack was also found to be successful in exploiting the privilege escalation mechanism of NVIDIA's GPUs, allowing the attackers to gain host root access. The researchers used a combination of techniques, including page-table privilege escalation and exploiting the triple-bit SDC, to achieve this. The attack was also found to be successful in exploiting the single-error-correct, double-error-detect (SECDED) ECC on these GPUs, which resulted in silent data corruption (SDC).
The researchers behind the GPUThor attack advised that the attack does not carry a CVE identifier and that no in-the-wild exploitation has been reported as of August 27, 2026. However, they also noted that the attack code is due to be made public on November 15, 2026, which may allow other researchers to explore and exploit the attack.
In response to the discovery of the GPUThor attack, NVIDIA released a security notice with guidance on how to mitigate the attack. The notice advised that ECC remains a sufficient mitigation, but it can no longer be treated as a sufficient defense. NVIDIA also advised that a complete fix would require stronger multi-bit error correction and in-DRAM defenses, such as Refresh Management or Per-Row Activation Counting, in future GPUs.
In conclusion, the emergence of the GPUThor attack has highlighted the limitations of NVIDIA's error correction mechanism in their high-end GPUs. The attack has significant implications for the security and stability of NVIDIA's high-end GPUs and highlights the need for stronger multi-bit error correction and in-DRAM defenses.
Related Information:
https://www.ethicalhackingnews.com/articles/New-GPUThor-Rowhammer-Attack-Reveals-Limitations-of-NVIDIAs-Error-Correction-Mechanism-ehn.shtml
https://thehackernews.com/2026/08/gputhor-rowhammer-defeats-ecc-on-nvidia.html
Published: Sat Aug 29 22:05:23 2026 by llama3.2 3B Q4_K_M
