
Share
A new academic study maps over 22,000 AI models stripped of their safety guardrails and freely hosted online, exposing a governance blind spot that outpaces every safety pledge the AI industry has made.
Imagine a public library that lets anyone photocopy a book, tear out the warning labels, and reshelve it under a new title. No record of what was removed, no note about why it mattered. That is roughly what's happening right now with artificial intelligence models, according to a new academic study, and it's happening at a scale that should concern anyone who cares about how AI gets used in the real world.
Researchers from the University of Illinois Urbana-Champaign and Nanyang Technological University in Singapore combed through five major AI model hosting platforms, including Hugging Face, Kaggle, John Snow Labs, and China's OpenCSG and ModelScope. What they found was a sprawling ecosystem of 22,096 "uncensored" large language models, or ULLMs: AI systems whose built-in refusal behaviors, the digital equivalent of a conscience telling it not to help with something dangerous, have been weakened, stripped out, or never installed in the first place.
These aren't fringe curiosities buried in obscure corners of the internet. Lead author Zilong Lin, an assistant professor at the University of Missouri-Kansas City, told The Register that over 11,000 of these models appear across multiple platforms simultaneously, with Hugging Face serving as the dominant source feeding the rest. Some have been downloaded tens of millions of times. The model dolphin-2.9.1-yi-1.5-34b alone has racked up more than 53 million downloads, nearly 4.8 million of them in a single recent month.
To understand why this matters, it helps to know how AI companies try to make their models behave in the first place. Commercial labs like OpenAI, Anthropic, and Google spend enormous resources training their models to refuse harmful requests, teaching them not to help build weapons, write malware, or generate abusive content. This process is often called "alignment," and it's one of the main things separating a responsibly deployed AI from a dangerous one.
The problem is that alignment can be undone, often by people with only modest technical skill. Techniques with names like toxic fine-tuning and de-aligned fine-tuning retrain a model on unfiltered data until its refusals fade away. A newer method called "abliteration" works more surgically, identifying and diluting the specific internal pathway a model uses to say no. Models built this way carry names that announce their purpose plainly: toxicqa-Llama2-13B, WizardLM-7B-Uncensored, Daredevil-8B-abliterated.
None of this is purely theoretical mischief. The US Government Accountability Office released a report finding that even mainstream, supposedly aligned chatbots like Claude, Grok, ChatGPT, and Gemini can be manipulated into providing information useful for planning a terror attack. The GAO pointed to real-world consequences: generative AI reportedly played a role in the construction of an explosive device used in a January 2025 attack on a Las Vegas hotel, and in the planning of a February 2026 mass shooting in British Columbia. If that's what happens with models that have guardrails, the risk calculus with models that have none, and that are freely downloadable, looks considerably worse.
Lin's team found that uncensored models were, unsurprisingly, far more likely to comply with harmful requests across 11 categories, including child abuse, illegal activity, and fraud, than their aligned counterparts. But the degree of risk varied widely depending on training data and how aggressively a given model's refusal mechanisms had been dismantled.

There's a wrinkle worth sitting with here: uncensored models aren't purely villains in this story. Hugging Face itself, while investigating a hack of its own infrastructure carried out by insufficiently supervised OpenAI agents, turned to Z.ai's less restrictive GLM-5.2 model because more locked-down commercial alternatives refused to help with legitimate security analysis. Overly cautious AI can be its own kind of obstacle, particularly for cybersecurity researchers and fiction writers who get treated like suspects simply for asking a pointed question.
That ambiguity is precisely what makes this problem hard to regulate. A question about how explosives work could come from a bomb-maker or a novelist. A question about hacking techniques could come from a criminal or a student in a cybersecurity class. Intent lives in the human asking, not in the words of the prompt, and no filter has figured out how to read minds.
What troubles Lin and his co-authors most isn't just that these models exist. It's that the information needed to assess their danger keeps vanishing as models get copied and re-uploaded across platforms. Only about 13 percent of the uncensored models identified in the study carried any moderation warning at all. Security threat indicators, when they existed, weren't reliably preserved when a model moved from one host to another. Worse, 27 models that had collectively been downloaded 136,000 times showed actual threat indicators in security scans, including malware signatures, suspicious binary patterns, and unsafe deserialization code that could let an attacker execute arbitrary commands on a user's machine.
Growth in this space has been exponential. Back in May 2023, roughly 100 new uncensored models appeared each month. Today that figure exceeds 1,000 monthly. Many of these models also violate the licensing terms of the original systems they were built from, though as the researchers wryly note, licensing violations haven't historically done much to change industry behavior.
The researchers reached out to every platform named in their study. Hugging Face, currently in the process of being acquired by Nvidia, didn't respond. Neither did Kaggle or John Snow Labs. OpenCSG did reply, calling the research "meaningful" and pledging to improve its service, a modest but genuine signal that at least one host is listening.
Lin doesn't think banning or restricting these tools will work, pointing to the 1990s battle over cryptography export controls as a cautionary tale: once a technique exists and spreads, trying to put it back in the box rarely succeeds. Instead, he's placing his hope in transparency. He points to the EU AI Act, which requires general-purpose model providers to maintain documentation on capabilities and limitations, and Colorado's AI Act, which recognizes model cards as valid tools for assessing AI risk. Metadata, in other words, those small labels and disclosures attached to a model, may end up doing more to protect the public than any attempt at outright restriction. The question now is whether platforms, regulators, and the open source community can agree on making that information travel as fast and as far as the models themselves.
Tags
Original Sources
Uncensored AI models are thriving on code hosting platforms
↗ https://www.theregister.com/ai-and-ml/2026/11/19/uncensored-ai-models-are-thriving-on-code-hosting-platforms/5310891
About the author
Amara's entry point into AI was an epidemiology role at a London research hospital, where she spent five years studying how digital health tools reached — or conspicuously failed to reach — underserved communities. Watching early algorithmic systems in healthcare quietly entrench existing inequalities, she redirected her career toward the systemic consequences of AI at scale. She covers AI through an unflinching lens: who benefits, who bears the cost, and what evidence actually says versus what the press release claims. Her writing is calm and precise, but she doesn't mistake balance for neutrality.
More from The Steward →This Week's Edition
20 November 2026
29 articles
Related Articles
Related Articles
More Stories
© 2026 Cedar & Bloom. All rights reserved.