Hugging Face malware campaign returns to hijack GPUs for crypto mining

A crypto mining and backdoor campaign using Hugging Face model repositories has resurfaced after being taken down. The most popular malicious model was downloaded over 28,000 times.

In addition, researchers have identified a malicious Solana starter kit that has been active on GitHub since late 2025 that is designed to deceive AI code writers into deploying a remote access Trojan. Researchers at Pluto Security identified 31 malicious repositories hosted on 7 Hugging Face accounts.

The models are disguised as low scale Qwen2 checkpoints and are designed to attack GPU hosts using self-hosting inferencing servers including Xinference and GPUStack. With models linked they have been downloaded over 32,000 times. This does not mean they were successfully infected, says Pluto.

Rather than using poisoned weights or pickle exploits, the attack abuses the trust_remote_code flag in Hugging Face which allows models to include their own Python code. When enabled this code is executed on the users machine as soon as they load the model read more about Hugging Face malware campaign returns to hijack GPUs for crypto mining

Get up to date on the latest cybersecurity news and enhance your knowledge of cybersecurity with our thorough coverage of the dangers, breaches, and solutions.

Leave a Reply

Your email address will not be published. Required fields are marked *