On October 29, Pluto Security published an analysis of the Load/Mine campaign, in which Hugging Face model repositories delivered cryptocurrency miners and code that gave attackers remote access. The researchers identified 31 malicious repositories across seven accounts. After the first wave was removed, new accounts appeared.
How the infection works
The attack abuses the ability to run Python code that ships with a model. Enabling trust_remote_code allows that code to execute while the model loads. This means the malicious instructions can run before anyone actually starts using the model.
The repositories posed as small Qwen2 test models. The samples described in the report installed a crypto miner. Some also opened a connection that let the attacker run commands on the server. BleepingComputer confirmed that one copy of the model contained both the miner and this remote access mechanism.
This does not mean every model on the platform is dangerous. The risk comes from running unverified code supplied by a specific repository author.
Who is at risk
Teams that run models on their own GPU servers should be especially careful, including those using tools such as Xinference and GPUStack. How far a compromise can reach depends on the process’s privileges and the data it can access.
More than 32,000 downloads of the identified and related repositories does not mean the same number of infections. Downloading files and executing malicious code are two different events. The most popular of the identified models had more than 28,000 downloads.
What to do now
- Check the repositories you use against the indicators of compromise in the Pluto Security report.
- Find where
trust_remote_code is enabled and who approves that decision.
- Review model code before running it, and pin approved versions.
- Isolate any suspicious server, preserve evidence, and check for persistence mechanisms.
- If malicious code was executed, rotate any credentials accessible from that server.
SecurHUB’s take
Code that arrives with a model needs the same scrutiny as any other software dependency. Incident response should include determining which secrets and systems the process that loaded the model could reach.
Sources: Pluto Security, BleepingComputer