• Sources: primary
  • Summary: Pluto Security found 31 malicious repositories across seven Hugging Face accounts posing as small Qwen2 checkpoints aimed at Xinference and GPUStack hosts, and a second wave appeared within days of the first takedown. The code runs when trust_remote_code is enabled, installs a miner disguised as a system process, and in some versions opens a reverse shell or a container escape script, while Hugging Face's scanners passed it. Separately, Air Security reports a fake Solana starter kit on GitHub that ranks highly for AI coding agents and tells them to run a PowerShell one-liner that installs NetSupport RAT.
  • Why it matters: Inference platforms that load models with trust_remote_code run arbitrary Python, so treat it as code execution and restrict it to reviewed repositories.