> Back to All Posts

Hugging Face Malware Hijacks GPU Servers Days After Takedown

Hugging Face malware

Fake AI models now work as a delivery system for crypto mining code. Hugging Face malware came back within days of a takedown, under fresh accounts and new repos. One fake model pulled more than 28,000 downloads while scans kept marking it clean. This Hugging Face malware goes after the GPU server itself, so anyone running self-hosted AI on their own hardware has a reason to look.

The model file was mostly empty

Researchers at Pluto Security counted 31 bad repos across seven accounts, posted in two waves since September. The files posed as small Qwen2 test checkpoints, the kind people load to check a setup works and then forget. Counting linked repos, the haul passed 32,000 downloads. That number tracks pulls rather than real infections.

The top repo, MENXIANGYU/qwen2-tiny, held no working model at all. It carried Python that drops a miner dressed as a normal Linux process. The Hugging Face malware restarts itself when killed, and it kills rival miners so one operator keeps the whole machine. Some builds open a reverse shell too, handing the attacker a live connection to the box.

Why Hugging Face malware passed every scan

The trick sits in trust_remote_code, a built-in option that lets a model ship and run its own Python. Flip it on and that code fires the second the model loads. Many self-hosted tools ask for the setting by default. So it rarely gets a second thought.

Hugging Face screens uploads with antivirus plus tooling from Protect AI and JFrog. Those checks hunt for known signatures and unsafe model formats. What they skip is the question of what a model’s own Python does once it runs, and Hugging Face malware walks straight through that gap. So the lead model of the second wave scanned clean while its download count climbed past 28,000.

GPU servers make it worth the effort

The fake checkpoints went after hosts running self-hosted inference servers such as Xinference and GPUStack. Those machines pack real compute and run around the clock. Their owners expect heavy GPU load anyway. Mining hides in plain sight, which is why Hugging Face malware aims at research boxes rather than laptops.

At one point the team watched nine machines start mining at once, all of them apparently one shared research cluster. The jobs stopped six minutes later, which fits an admin spotting the load and killing them. Later payloads also grab cloud metadata and GPU details, while one script can break out of a container to reach the host underneath.

The takedown bought a few days

Pluto found its way in through a sloppy test account, where backup files left behind pointed at the attacker’s main profiles. Hugging Face and China’s ModelScope both pulled the reported accounts. A second wave of 19 repos then turned up under new accounts within days of the first removals.

Hours after that report landed, the operator posted another model carrying a container escape script. Second-wave repos stayed live for eight days after the flag went in. Downloads on the lead model climbed from 7,354 to 28,412 by 28 October. A near-identical copy posted on 25 October under the account htt0118 kept the same miner and the same reverse shell, which tells you how little a single takedown slows this Hugging Face malware down.

Small payouts, stubborn persistence

Blockchain records put the operator’s main wallet at roughly $5,000 in Pearl (PRL). That looks thin. But the wallet kept drawing from mining pools after the repos vanished, and stolen compute costs the attacker nothing.

Pluto ties the campaign to a money-driven Chinese-speaking operator. The evidence covers code comments in simplified Chinese, command and control servers on Alibaba Cloud, and work hours matching late nights in China. None of it names a person.

AI coding agents face a similar trap

Separate research describes a GitHub repo, SolanaUpdateTools/solana-starter-kit, that comes up as a top result when AI coding agents hunt for a Solana template. The repo carries 286 stars and dates to November 2025. Its README is the only file there, and the app code it promises does not exist. Researchers connected it to a NetSupport operator already on record, plus an earlier scheme that baited crypto traders with fake TradingView signal tools.

The single install step tells you to paste one Windows PowerShell line into a terminal. That command builds a URL from fragments, then runs a script in memory that installs NetSupport Manager as a remote access tool. It then prints a fake error about developer features pausing for 72 hours. An agent would pass that along as a routine server hiccup.

What to do if you loaded one of these models

Treat any machine that opened one of these repos with trust_remote_code enabled as compromised. Killing the miner fixes nothing, because the Hugging Face malware leaves SSH keys, scheduled tasks and edited startup files behind. Rotate every credential the attacker could have reached. Then check the host against the indicators in the published research.

Leave trust_remote_code off unless you have read the repo’s Python yourself, and keep inference servers off the open internet by reaching them over a private tunnel or a VPN instead of exposed ports. Watch GPU use and outbound traffic too. A miner that dodges file scanners still has to phone a pool.

Final Thoughts

Hugging Face malware has moved the target from the model file to the step that loads it. Scanners were never built for that. Model repos now carry the same trust problem that package registries have fought for years, except the payload lands on hardware worth thousands a month.

Download counts in the tens of thousands suggest plenty of people skipped the part where someone reads the code. Until platforms close that gap, treat any model asking for remote code execution like an unknown binary. Handle the convenience setting as the risk it is, and the Hugging Face malware problem shrinks fast.

Janet Andersen

Janet is an experienced content creator with a strong focus on cybersecurity and online privacy. With extensive experience in the field, she’s passionate about crafting in-depth reviews and guides that help readers make informed decisions about digital security tools. When she’s not managing the site, she loves staying on top of the latest trends in the digital world.