Work

Memory Safety for Commodity Graphics Processing Units

Year: 2026
Type: conference-paper
Abstract: GPUs accelerate a growing fraction of modern computing workloads. Yet, modern GPU programming frameworks, including CUDA, ROCm, and Triton, expose unchecked pointer arithmetic, leaving GPU kernels vulnerable to the same spatial memory errors as C and C++ programs. Despite prior work studying the exploitation of memory errors in GPU applications, GPU vendors and developers have not been convinced and consequently continue to largely treat such errors as reliability concerns rather than security issues, leaving vulnerabilities unpatched and potentially exploitable. During the first two years of this PhD, we have challenged this assumption. In particular, we have demonstrated that remote, unprivileged adversaries can trigger device-side memory errors and, under specific conditions, escalate them into write-what-where primitives, enabling denial-of-service, targeted manipulation and degradation of ML models, sensitive data leakage, and device-side code injection. At the same time, we notice that existing GPU memory-safety tools have seen limited adoption as they impose high overhead, require hardware changes, or fail to integrate with production GPU software stacks. Therefore, we propose three lightweight memory-safety mitigations, based on insights from our offensive research, tailored to the strengths of GPUs and their deployments. Specifically, we discuss continuous fine-grained memory-layout re-randomization, data-space randomization, and Right-on-Time Safety (RoTS): a software-only memory-safety mitigation that validates GPU memory-access ranges at the host-side kernel launch whenever those ranges depend only on launch-time values, and emits device-side checks only for data-dependent accesses.
Author: Jonas Roels
Institution: KU Leuven
Language: English
Open Access status: hybrid
OpenAlex ID: W7225900065
FWCI:
Cites: 21
Cited by:
Related to: 10