Nvidia GPUs Vulnerable to New Memory Bit Flipping Attack: GPUThor

Researchers from the University of Toronto have developed a new memory bit flipping technique called GPUThor that can defeat Nvidia's error-correcting codes (ECC) defense and lead to root access on the underlying system. This technique, dubbed Rowhammer, exploits the cell density of modern random access memory (RAM) chips and can cause significant security implications, including privilege escalation and AI model manipulation.

GPUThor is the first Rowhammer attack on Nvidia GPUs to break through ECC, a defense mechanism that was previously thought to be secure. The technique works by intentionally triggering bit flips using rapid and repeated read operations on the same memory row, which can have devastating consequences for enterprise and server-class GPUs.

The researchers tested GPUThor on four different Nvidia Ampere GPUs with GDDR6 memory and found that it could produce double- and even triple-bit errors, which the built-in ECC mechanism was not designed to handle. This made it possible for the attackers to escalate their privileges to root level and access sensitive data.

While the researchers did find some limitations to GPUThor, they were able to crash GPUs so often that within one day the cards flagged themselves as defective and due for replacement using their internal crash detection mechanisms. This highlights the importance of monitoring Nvidia's error-correction counters, as a spike in these counters could be an indication that an attack is in progress.