New GPU Memory Attacks Let Hackers Seize Full Control of Nvidia Machines

Security researchers have developed three new hardware-level attacks — GDDRHammer, GeForge, and GPUBreach — that exploit Rowhammer-style memory vulnerabilities in Nvidia GPU hardware to seize complete control of affected machines, including their CPUs.

Rowhammer attacks work by repeatedly accessing specific rows of memory cells to induce bit-flips in adjacent rows, corrupting data in ways that can be weaponised to escalate privileges or hijack system processes. First demonstrated against DRAM in 2014, the technique has affected PC and server security for over a decade. These new attacks extend that threat surface dramatically — from system RAM into GDDR (Graphics Double Data Rate) memory, the high-bandwidth memory that powers modern GPU hardware.

How GPU Memory Becomes a Gateway to Full System Compromise

The three attack variants each approach the problem differently, but share a common outcome: by inducing targeted bit-flips in GPU memory, attackers can corrupt data structures that the CPU trusts, ultimately gaining kernel-level or hypervisor-level control of the entire machine. GPUs and CPUs increasingly share memory spaces and communicate through high-trust interfaces — a design optimised for performance in AI workloads that, according to the researchers, also creates exploitable pathways.

By hammering GDDR memory, attackers can corrupt data structures the CPU trusts — turning a graphics card into a skeleton key for the entire system.

The attacks do not require physical access to the target machine. In cloud environments — where multiple tenants share physical servers equipped with high-end Nvidia GPUs — the implications are particularly serious. A malicious cloud tenant could, in theory, use these techniques to break out of their isolated environment and access another customer's data or processes.

Who Is at Risk, and Why AI Infrastructure Is the Focal Point

The risk is concentrated wherever Nvidia GPUs are deployed at scale — which, in 2026, means a substantial portion of the world's AI training and inference infrastructure. Data centres running Nvidia H100 and A100 accelerators for large language model training represent high-value targets. So do cloud platforms offering GPU instances to researchers, startups, and enterprises.

For individual consumers, the immediate risk is lower, though machines running Nvidia consumer GPUs are not categorically immune. The attack complexity and the need for local or remote code execution as a precondition mean that casual users face a smaller — though not zero — threat profile.

The human impact extends beyond individual machines. AI systems trained or operated on compromised infrastructure could be manipulated at the hardware level, below the visibility of conventional software security tools. Model weights, training data, and inference outputs could all be tampered with in ways that are extremely difficult to detect after the fact.

What Nvidia and the Research Community Are Saying

As of publication, Nvidia had not issued a public patch or formal security advisory in response to the disclosed vulnerabilities, according to Ars Technica's reporting. The attacks were disclosed through the research community, and the timeline and terms of any coordinated disclosure with Nvidia were not specified in initial reporting.

Rowhammer defences have historically proven difficult to implement without significant performance trade-offs. Target Row Refresh (TRR), a mitigation built into modern DRAM, was itself bypassed by researchers in 2020 with a technique called TRRespass. Whether analogous mitigations exist or can be developed for GDDR memory — which is architecturally distinct from DRAM — remains an active area of concern.

Hardware-level attacks of this class are notoriously difficult to patch after the fact. Unlike software vulnerabilities, which can be addressed through updates, Rowhammer-style flaws are rooted in the physical properties of memory cells. Mitigations tend to come in the form of firmware changes, memory controller updates, or — in extreme cases — hardware redesigns, none of which can be deployed quickly at scale.

The Broader Shift: GPUs as Critical Security Infrastructure

For years, GPU security was treated as a secondary concern — graphics cards were peripherals, not attack surfaces. That assumption has changed under the weight of the AI boom. Nvidia GPUs now sit at the heart of the most sensitive computational workloads on the planet, from national security applications to financial modelling to medical AI.

The emergence of GDDRHammer, GeForge, and GPUBreach signals that the security research community — and by extension, the threat actor community — has turned serious attention to GPU hardware as a target. Cloud providers, AI companies, and enterprise IT teams that have invested heavily in GPU infrastructure will need to reassess their threat models to account for hardware-layer attacks that existing security tooling is not designed to detect.

What This Means

Any organisation running Nvidia GPUs in a multi-tenant or networked environment should treat this disclosure as an urgent prompt to review hardware security posture and monitor for guidance from Nvidia — because a full software patch for a hardware-rooted vulnerability may not be possible.