Black Hat USA 2026 | GPUBreach: Privilege Escalation Attacks on GPUs Using Rowhammer
Rowhammer attacks have been extensively studied on CPUs, where they have enabled powerful exploits, including privilege escalation. In contrast, Rowhammer on NVIDIA GPUs, despite recently demonstrated by GPUHammer attacks, has largely been viewed as low impact, limited to inducing random bit flips that merely degrade machine learning accuracy. In this Briefing, we will overturn that assumption and show that GPU Rowhammer can be weaponized into a full-system compromise.
We will demonstrate GPUBreach, the first targeted Rowhammer attacks on NVIDIA GPUs that enable privilege escalation and cross-process memory access. By reverse engineering GPU page table allocation behavior, we identify when and where page tables are allocated, allowing an unprivileged CUDA kernel to induce deterministic bit flips in victim GPU page table entries. This enables corruption of GPU page tables and yields arbitrary read and write access to GPU memory across isolation boundaries. We use these primitives to demonstrate concrete attacks, including the extraction of cryptographic keys from GPU-accelerated libraries like cuPQC, the exfiltration of proprietary machine learning model weights, and the stealthy modification of GPU kernel code (e.g., in cuBLAS) to universally degrade the accuracy of any ML model.
Most critically, we show that GPU compromise does not stay contained. We bridge the GPU-CPU boundary to achieve full system takeover despite IOMMU protections that are designed to protect the host against a malicious device. Leveraging GPU-side arbitrary writes to inject malicious DMA operations to the CPU memory, we exploit a previously unknown memory safety vulnerability in the NVIDIA GPU driver stack, and ultimately gain arbitrary write access in the Linux kernel, resulting in root shell access.
These results fundamentally challenge the security assumptions of modern heterogeneous systems and demonstrate that GPUs are not just an accelerator, but powerful attack surfaces that break the security of the entire system.
Takeaway
1. **Turning GPU Rowhammer into a targeted exploitation primitive capable of privilege escalation:** Attendees will learn how seemingly random GPU Rowhammer bit flips can be engineered into precise, attacker-controlled faults. We will walk through how to reverse engineer GPU page table structures and allocation behavior on NVIDIA GPUs, identify when and where page tables are placed in memory, and use memory “massaging” techniques to position them in vulnerable regions. These insights enable deterministic corruption of page table entries, yielding arbitrary read/write access to GPU memory and breaking isolation across processes and containers.
2. **Exploiting New NVIDIA GPU Driver Memory Safety Vulnerabilities.** Attendees will gain a practical understanding of how GPU compromise can be extended to the host. We will show how GPU-controlled DMA can be used to tamper with privileged data structures in the NVIDIA Driver on the host. Through a deep-dive of the CPU-GPU communication buffers in the open-source NVIDIA GPU Driver, we will show how trusted-but-unsanitized inputs from the device side lead to out-of-bounds writes within the drivers, and how this can be turned into powerful primitives such as arbitrary kernel writes and function pointer overwrites within the Linux kernel, that can lead to system-wide privilege escalation.
3. **Rethinking GPU Security Assumptions:** This talk encourages attendees to rethink traditional GPU trust models. For GPU cloud operators and users, this highlights the immediate need to adopt recommended Rowhammer mitigations such as enabling ECC on all GPUs. For Linux kernel and device driver developers, this encourages a need to rethink the trust-assumption in GPUs, harden the drivers against a new threat of “malicious” GPUs. We will discuss practical defensive strategies, ranging from hardware and firmware mitigations against Rowhammer-induced page-table corruption to software techniques for hardening drivers and systems against compromised or malicious GPU devices.
These insights help security practitioners understand the real attack surface of GPUs and evaluate risks in AI infrastructure, workstations, and shared GPU environments.
Chris S. Lin | PhD Student, University of Toronto
Yuqin Yan | PhD Student, University of Toronto
Guozhen Ding | Master Student, University of Toronto
Joyce Qu | Undergraduate Student, University of Toronto
Joseph Zhu | Machine Learning Engineer, Google
David Lie | Professor and Canada Research Chair, University of Toronto
Gururaj Saileshwar | Assistant Professor, University of Toronto
https://blackhat.com/us-26/briefings/schedule/?#gpubreach-privilege-escalation-attacks-on-gpus-using-rowhammer-52286Black HatRead More