Hugging Face malware campaign returns to hijack GPUs for crypto mining
Pluto Security identified 31 malicious Hugging Face model repositories across seven accounts. The repositories posed as small Qwen2 test checkpoints and targeted self-hosted inference servers on GPU hosts, including Xinference and GPUStack. When enabled, trust_remote_code runs model-supplied Python code on the user's machine as the model loads. The code installed a cryptocurrency miner, restarted it, and in some variants opened a reverse shell. New repositories appeared within days after accounts were removed. The source also describes a separate fake Solana starter kit on GitHub that could trick AI coding agents into installing NetSupport Manager.
CERT24