GPU → Vulnerability

GPU vulnerability: 8 security risks and how to address them

GPUs have revolutionized everything from AI training to high-performance computing, but their rise in server environments has also introduced a new frontier of security risks. While you’re likely well-versed in CUDA cores, memory bandwidth, and parallel processing, have you considered how these same architectural strengths can become attack surfaces?

Unlike traditional CPU servers, GPU-powered systems face unique threats—from side-channel exploits to virtualization vulnerabilities—that can compromise performance, data integrity, and even entire infrastructures. If you think your GPU workloads are safe just because they’re running in a secure data center, it’s time to take a closer look.

Get premium GPU server hosting

Unlock unparalleled performance with leading-edge GPU hosting services.

1. Side-channel attacks

Side-channel attacks exploit unintended information leakage from hardware, such as timing variations, power consumption, or electromagnetic emissions. While side-channel attacks are common in CPUs, GPUs are particularly vulnerable due to their parallel processing nature. Since multiple processes may share the same GPU hardware, an attacker running code on the same GPU can infer secret data by observing execution patterns, memory access times, or cache behavior.

One well-known GPU side-channel attack is keystroke inference, where an attacker monitors GPU-based rendering workloads to extract keystrokes or other user input patterns. Similarly, researchers have shown that cryptographic operations processed on a GPU can be susceptible to timing attacks, revealing secret keys.

Mitigation strategies:

2. GPU memory leakage

Unlike CPUs, GPUs do not always have robust memory isolation, especially in multitenant  environments. If memory is not properly cleared when a process ends, an attacker could retrieve leftover data from another user’s workload. This issue is particularly concerning in cloud GPU environments where different customers use the same physical hardware.

For example, in some NVIDIA CUDA and AMD ROCm implementations, improperly deallocated memory can retain portions of deep learning model weights, sensitive images, or cryptographic data. This could allow a malicious user to extract data from previous workloads, potentially revealing proprietary or confidential information.

Mitigation strategies:

3. Driver and firmware exploits

GPU drivers and firmware are frequently updated to patch security vulnerabilities, but outdated or improperly configured drivers present an easy attack vector. Exploits in GPU drivers can enable attackers to execute arbitrary code at the kernel level, escalate privileges, or crash entire systems.

For instance, past vulnerabilities in NVIDIA and AMD drivers have enabled remote code execution (RCE) by sending malicious shader code or malformed API calls. Attackers can also use privilege escalation exploits to gain admin control over GPU-accelerated servers, compromising the entire system.

Mitigation strategies:

4. GPU-specific malware

Traditional antivirus and endpoint protection tools focus on detecting CPU-based threats, leaving GPUs relatively unprotected.

GPU-specific malware can execute malicious code directly within the GPU’s memory, making it harder to detect with standard security tools. Some GPU rootkits, such as Jellyfish (a proof-of-concept GPU malware), have demonstrated the ability to persist in GPU memory and avoid detection by conventional security software.

Attackers can also use GPUs for malware obfuscation, offloading key parts of an exploit to the GPU to evade detection. Since GPU execution operates separately from the CPU, traditional forensic tools may not even detect the presence of malicious code.

Mitigation strategies:

5. Virtualization security risks

GPU virtualization allows multiple virtual machines (VMs) to share a single GPU, which introduces risks such as GPU escape attacks—where an attacker in one VM gains access to another VM’s memory or workloads. Vulnerabilities in NVIDIA vGPU, AMD MxGPU, or Intel GVT-g have led to cases where guest VMs could crash the host system or escalate privileges.

One major concern is GPU passthrough attacks, where an attacker gains direct access to a GPU in a VM and exploits vulnerabilities to bypass hypervisor security. For example, attackers could modify a VM’s GPU memory mappings to access another tenant’s data or even the host system.

Mitigation strategies:

6. AI and machine learning poisoning attacks

Since many GPU servers are used for AI training, they are susceptible to data poisoning attacks, where an attacker injects manipulated data into the training pipeline to alter the model’s behavior. This can lead to biased predictions, backdoor vulnerabilities, or security loopholes in deployed AI applications.

For instance, an attacker could introduce poisoned images into a facial recognition dataset to cause misclassification, or insert adversarial examples into a fraud detection model to reduce its effectiveness.

Mitigation strategies:

7. Cryptojacking and GPU resource abuse

Cryptojacking is when an attacker hijacks GPU resources to mine cryptocurrency without the owner’s consent. Cloud-hosted GPU servers are particularly vulnerable, because attackers can scan for instances with weak security and deploy cryptojacking scripts.

One common method involves compromising exposed SSH services or exploiting vulnerabilities in containerized environments. Since cryptojacking malware often runs at full GPU capacity, it can degrade server performance and increase energy costs.

Mitigation strategies:

8. Buffer overflows

Buffer overflow vulnerabilities occur when a program writes more data to a buffer (temporary memory storage) than it can hold, leading to memory corruption. This can allow attackers to overwrite adjacent memory regions, execute arbitrary code, or crash a system.

While buffer overflows are a well-known issue in CPU-based applications, they are particularly dangerous in GPUs due to the complexity of GPU drivers, shader compilers, and parallel processing frameworks.

GPU buffer overflows can arise from several sources:

In 2022, for example, security researchers identified a buffer overflow vulnerability in NVIDIA’s CUDA driver, allowing local attackers to escalate privileges and execute arbitrary code. Similar vulnerabilities have been found in GPU shader compilers, where specially crafted shader programs could trigger overflows to gain unauthorized access.

Mitigation strategies:

Additional GPU security strategies

To pre-emptively secure GPU servers, consider:

While GPU servers share some security risks with traditional dedicated servers, their specialized processing and shared multi-user environments introduce additional attack surfaces that need to be addressed.

Additional resources

What is a GPU? →

A complete beginner’s guide to GPUs and GPU hosting

Best GPU server hosting [2025] →

Top 4 GPU hosting providers side-by-side so you can decide which is best for you

A100 vs H100 vs L40S →

A simple side-by-side comparison of different NVIDIA GPUs and how to decide

Amy Moruzzi is a Systems Engineer at Liquid Web with years of experience maintaining large fleets of servers in a wide variety of areas—including system management, deployment, maintenance, clustering, virtualization, and application level support. She specializes in Linux, but has experience working across the entire stack. Amy also enjoys creating software and tools to automate processes and make customers’ lives easier.