Malicious AI Models on Hugging Face Spread Cryptomining Malware and Backdoor
November 2, 2026 · 1 min read
A security investigation uncovered malicious artificial intelligence models on the Hugging Face platform that embed cryptocurrency‑mining malware and a remote‑access backdoor. Researchers identified 31 harmful model repositories that were uploaded through seven different Hugging Face accounts, and the total number of downloads surpassed 32,000.
Pluto Security reported that the attackers disguised the malicious repositories as ordinary small AI models. When users load a model with the option trust_remote_code=True, the embedded Python code executes, installing a cryptomining program that uses GPU resources and a backdoor that allows remote control of the infected system. The malware masquerades as a legitimate system process to evade detection and is designed to restart the mining operation even after it is stopped. Some variants also enable attackers to issue remote commands on compromised servers.
Although the exact number of infected systems and the financial damage remain unknown, the incident highlights the risk that public AI models pose to corporate GPU servers. Security experts advise AI development firms and research institutions to verify the source and Python code of any external model before adoption, to restrict unnecessary remote code execution, and to avoid enabling trust_remote_code=True. Unverified models should be run in isolated environments, and organizations should monitor for abnormal GPU usage and suspicious communications with external servers.
The findings underscore the need for stricter vetting of community‑shared AI resources to prevent similar supply‑chain attacks.