Skip to content
Security & Trust Talking point

An attacker with GPU memory access could copy or quietly corrupt an AI model

Black Hat researchers showed that, once inside GPU memory, an attacker could steal model weights or alter model code.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
An attacker with GPU memory access could copy or quietly corrupt an AI model
In brief
  • Black Hat USA 2026 researchers showed that an attacker with GPU memory access could copy a model's weights and change the code that runs it.
  • One speaker said even a legitimate admin cannot easily inspect GPU code while it is loaded. The researchers did inspect it, using NVIDIA's cuobjdump tool.

At Black Hat USA 2026, researchers argued that an attacker with read and write access to a GPU's memory could steal an AI model or quietly damage it. The talk, called GPUBreach, was given on the Black Hat show. The transcript does not name the speakers.

What was said

The talk opened with a puzzle. A speaker described a system where the model runner looked current and the OpenVAS scanner found nothing wrong. The question was how an attack could still happen. OpenVAS is a tool that scans for known vulnerabilities.

The answer ran through the graphics chip, or GPU. A speaker said a model sits in GPU memory for a long time while it waits for requests. Dumping that memory to steal the weights is "not very difficult," they said. The weights are the numbers that make up a trained model. The team said it could identify model families such as Llama 2, Llama 3, Mistral and Gemma from a dump.

The speakers called sabotage the more dangerous half. They described changing the GPU's own code, for example inside cuBLAS, an NVIDIA library for matrix maths. The change targets instructions that lower accuracy across the board but barely affect speed. That leaves little clue for debugging.

One speaker said "even the legit admin cannot easily inspect the code on the GPU while it's loaded." The researchers themselves did inspect loaded code, using NVIDIA's cuobjdump tool and their own read access.

Why it matters

Our reading: a clean scan may say little about whether a model's output can be trusted. The talk pointed at weights and maths code on the GPU, a place where a speaker said even administrators have trouble looking.

For leaders who buy or run AI, the question becomes who can inspect what runs on the GPU. A second question follows: how would anyone know it had changed? That judging falls to people. They must decide how far to trust a part they cannot easily see into.

The other side

The attacks need someone who already has arbitrary read and write access to GPU memory. The talk reached that point through Rowhammer, a way of flipping memory bits. It is a research demonstration, and the excerpts describe no real-world theft or tampering. The speakers also did not compare how easy the attacks are with how easy they are to detect.

Their advice was to "sanitize everything the GPU sends to CPU and treat the GPU as a non-trusted device." That advice protects the host computer from a malicious GPU. The excerpts give no remedy for stolen weights or tampered code.

The speakers also noted that modern systems have had a protection called IOMMU for more than five years. The excerpts end before the talk explains how the attack deals with it.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from