GPUBreach 攻撃について

A newly discovered vulnerability dubbed “GPUBreach” demonstrates that GPU-based Rowhammer attacks can now achieve complete system compromise. Scheduled for presentation at the IEEE Symposium on Security & Privacy in 2026, University of Toronto researchers revealed how manipulating GPU memory can lead to a full CPU root shell.

Most alarmingly, this exploit successfully bypasses standard hardware protections, representing a significant escalation in hardware-based threats. GPUBreach 攻撃は、従来のハードウェア保護を回避し、ハードウェアベースの脅威が大幅に進化したことを示しています。

攻撃の展開

Previously, GPU Rowhammer attacks primarily caused localized data corruption, such as slightly degrading the accuracy of machine learning models. GPUBreach elevates this threat by specifically targeting GPU page tables located within GDDR6 memory.

• The researchers reverse-engineered the NVIDIA driver to understand how these page tables are allocated and stored alongside standard user data.
• To execute the attack, threat actors utilize a timing side channel on Unified Virtual Memory (UVM) allocations to detect memory evictions. This exposes exactly when and where a new page table region is created.

By strategically freeing and filling memory, attackers can force GPU page tables to land immediately adjacent to vulnerable memory rows. A targeted Rowhammer bit-flip then alters a page-table entry, granting an unprivileged CUDA kernel arbitrary read and write access across all GPU memory.

攻撃の影響

The research team validated GPUBreach on an NVIDIA RTX A6000 graphics card, confirming several critical security impacts:

• Attackers can extract secret cryptographic keys directly from the NVIDIA cuPQC post-quantum library during active key exchanges.
• Malicious actors can stealthily plummet an AI model’s accuracy from 80% to zero by modifying a single code branch.
• Threat actors can scrape and steal highly sensitive Large Language Model (LLM) weights residing in GPU DRAM.

The exploit chains GPU memory access into a full CPU root shell by triggering newly uncovered memory-safety bugs in the NVIDIA kernel driver. This allows attackers to gain complete control over the system without being detected by standard hardware security mechanisms.

元記事: https://gbhackers.com/gpubreach-attack-could-lead-to-full-system-takeover/