TabSphere journal / Technology news
Malicious AI models abused to run cryptominers on GPU servers
Researchers found dozens of malicious model repositories that exploit a model-loading option to run Python code, install cryptominers and in some cases open backdoors. One fake model was downloaded over 28,000 times and,

This story is developing and details may change. Last checked 30 Oct 2026.
What happened
Reported by Hugging Face malware campaign returns to hijack GPUs for crypto mining, BleepingComputer on 29 Oct 2026. The event itself dates from 25 Oct 2026.
Security researchers reported a campaign that places malicious models in public model repositories. The malware uses a model-loading option to execute Python code when a model is loaded. The code installs a cryptocurrency miner disguised as a system process, removes competing miners, persists by restarting if killed, and some variants open a reverse shell or backdoor. Researchers reported dozens of malicious repositories published in waves, with many re-uploads after takedown. The most-downloaded malicious model and near-identical copies exceeded twenty-eight thousand downloads before removal. Researchers also observed mining activity on a shared research GPU cluster while investigating the campaign. These findings follow a related incident documented earlier by the repository host. 1 2
Why it matters to small businesses
If you or a supplier loads models from public repositories into self-hosted GPU servers, research clusters or inference setups that allow remote code execution at model load time, you could inadvertently run untrusted code on your machines. That code can run a cryptominer, use GPU capacity without your notice, remove competing processes, persist through restarts and in some cases provide a backdoor for further access. A single malicious model attracted very large download counts, showing how quickly an unsafe model can spread. The campaign also shows that takedowns can be followed by rapid re-uploads, so a single removal is not a guarantee of safety. For small businesses this matters where you host models or use third-party systems that run models with options that permit bundled code to execute. 1 2
TabSphere's view
This section is TabSphere's interpretation, not reported fact.
Interpretation: This is primarily a risk to organisations that run models on their own GPU infrastructure or use tools that automatically load models with permissive execution settings. Many small firms never host GPUs; for them the risk is low. But if you do run inference on-premises, or allow contractors or automated agents to fetch and run models without vetting, you should treat model repositories like software distribution channels and apply the same caution you use for packages and scripts. Prefer curated, audited models or provider-hosted inference where execution context is controlled.
What to do next
- Check how models are loaded. Review any workflow that fetches and loads models into servers. Identify settings or flags that allow downloaded model code to run at load time and disable those options unless you trust the source explicitly.
- Isolate GPU workloads. Run model inference on dedicated systems or containers separate from business-critical machines. Limit network access and credentials available to those hosts so a compromised model cannot reach other systems or services.
- Use vetted models and provider inference. Prefer models from curated hubs, official vendor models, or hosted inference APIs where the provider controls execution. Treat community uploads as unvetted software and require additional checks before use.
- Monitor resource and process behaviour. Watch GPU, CPU and network usage on hosts that run models. Unexpected sustained GPU-heavy work or unfamiliar system processes are warning signs that merit prompt investigation.
- Limit automated agent privileges. If you use AI coding assistants or automation that can fetch and run code or templates, restrict their permission to install or execute unreviewed packages or code on your servers.
Sensible precautions
- Avoid loading models directly from unknown or recently created repositories.
- Keep system packages and user accounts minimal on any GPU host.
- Back up important data and have a recovery plan for compromised hosts.
What isn't known yet
- Attribution of the campaign to a specific actor is tentative in the research; that remains an analyst assessment rather than established fact.
- Researchers observed re-uploads and additional malicious repositories, so the total scope and future persistence of the campaign may change as more activity appears.
This incident shows that machine-learning model repositories can be a source of malicious software just like any public code repository. If you run models on your own hardware, a few simple safeguards, checking how models are loaded, isolating workloads and watching resource use, will reduce your exposure without heavy cost.
If you want to review how your website, software or design interacts with third-party models or external code, we can have a calm, practical conversation about safer approaches.
Sources
Checked 30 Oct 2026.
