Malicious Hugging Face Models Resurface to Exploit GPU Resources for Crypto Mining

A new wave of malicious activity has emerged on the Hugging Face model repository, where attackers are distributing compromised AI models that commandeer GPU servers to run cryptocurrency miners. The campaign reappeared only days after an earlier takedown, indicating that the threat actors quickly adapted their tactics to continue siphoning compute power from unsuspecting hosts.

The illicit models are uploaded as seemingly legitimate machine‑learning assets, and one of them has been downloaded more than 28,000 times since its appearance. Researchers believe that the high download count reflects both the popularity of the platform and the reliance of developers on automated tools that fetch models without thorough verification.

Embedded within the model files is a payload that installs a backdoor and launches a mining process that targets the host’s graphics processing units. Because GPUs are the most efficient hardware for both AI inference and cryptographic hashing, the malicious code can generate significant revenue while remaining largely invisible to standard monitoring tools that focus on CPU usage.

Hugging Face hosts an open marketplace for thousands of pre‑trained models, a resource that has become integral to the rapid development of AI applications. This openness, while fostering collaboration, also creates a vector for abuse: malicious actors can masquerade as contributors, and the platform’s automated pipelines may inadvertently distribute harmful code to downstream users and AI‑driven coding assistants.

In a parallel development, a counterfeit Solana “starter kit” was posted on GitHub with the explicit aim of deceiving AI coding agents. The repository is crafted to look like a legitimate development bundle, but it contains scripts that, when executed by an automated assistant, pull in additional malicious components. Security analysts warn that such supply‑chain tricks could extend the reach of the mining campaign by compromising development environments.

The resurgence of the Hugging Face‑based mining operation underscores the need for stronger vetting processes on model‑hosting services and heightened awareness among developers who integrate external assets. Platform operators have pledged to tighten review mechanisms, while cybersecurity teams advise users to verify the provenance of models and to monitor GPU utilization for anomalous spikes. As attackers refine their methods, the community’s collective vigilance will be crucial in preventing further exploitation of AI infrastructure for illicit profit.