Skip to content
LIVE
The Executives BriefThe Executives Brief

22,096 uncensored AI models are loose online. Only 13% carry a warning.

Researchers mapped a model supply chain where the weights travel faster than the risk data attached to them, and 27 malware-flagged builds were downloaded 136,000 times.

ByOmar Al-BalawiTechnology Correspondent, The Executives Brief
·4 min read
22,096 uncensored AI models are loose online. Only 13% carry a warning.
Executive summary

Researchers Zilong Lin, Zichuan Li, Xiaojing Liao, and XiaoFeng Wang analyzed 22,096 uncensored large language models across Hugging Face, Kaggle, John Snow Labs, OpenCSG, and ModelScope, finding that only about 13 percent carried moderation warnings. For executives whose products route to third-party or open weight models, model provenance has just become a security, licensing, and liability question rather than a research curiosity.

Researchers scraped 22,096 uncensored large language models from five major hosting platforms and found that only about 13 percent of them carried any moderation warning. The paper, "Characterizing the Ecosystem of Open-Source Uncensored Large Language Models: Supply Chain, Governance, and Risks," comes from Zilong Lin, Zichuan Li, Xiaojing Liao, and XiaoFeng Wang, affiliated with the University of Illinois Urbana-Champaign and Nanyang Technological University in Singapore. Lin, the lead author and an assistant professor of computer science at the University of Missouri-Kansas City, told The Register that the platforms hosting these models, Hugging Face, Kaggle, and John Snow Labs in the US, plus OpenCSG and ModelScope in China, play a role similar to GitHub or GitLab in the open source world. The difference is that you cannot read a model file the way you read source code to see what it will do. Lin said evaluating whether a model has adequate safety safeguards typically requires deploying it and testing its responses to harmful prompts, which carries substantial computational and financial cost. That is why his team worked from publicly available model metadata instead.

Lin's team identified more than 11,000 uncensored models appearing across multiple platforms, with Hugging Face acting as the dominant upstream source. That movement creates what the authors call a governance gap: models travel more easily than the information needed to assess their risks. Security threat indicators were not consistently preserved across platforms, so a model can be copied from one host to another while the warnings attached to it are stripped away along the way. For anyone running a model routing service, an app, or a GitHub project that pulls weights from these hubs, the practical consequence is uncomfortable. Users of those apps or services may have no idea what the underlying model is, or what it was trained to refuse.

Producing one of these models does not require a lab. Techniques like toxic training or fine-tuning (TTFT) and de-aligned training or fine-tuning (DTFT) can be applied to datasets that were never safety filtered, yielding builds such as toxicqa-Llama2-13B and WizardLM-7B-Uncensored. A newer method, refusal-direction ablation, or abliteration, dilutes refusal training so the model is less likely to say no; Daredevil-8B-abliterated is one example. Lin said abliteration appeared in 2024 and quickly gained popularity within the open source LLM community, and within roughly two years a growing ecosystem of publicly available tools had emerged to support it. The volume tells the same story: about 100 uncensored models per month in May 2023, more than 1,000 per month presently. The model dolphin-2.9.1-yi-1.5-34b has been downloaded more than 53 million times, including more than 4.8 million last month alone.

Not all of this is a content problem. Twenty-seven of the models the researchers examined carried threat indicators from security scans, and those 27 have been downloaded 136,000 times. Many had signatures associated with malware, exhibited suspicious binary patterns, or contained unsafe deserialization mechanisms that could allow arbitrary code execution. That is a supply chain risk in the ordinary software sense, not just an alignment debate. The authors also note that many uncensored models violate the licensing terms of the models they were built on, though source code licensing has had little meaningful impact on the behavior of the AI industry.

The stakes are not hypothetical. The US Government Accountability Office just released a report finding that Claude, Grok, ChatGPT, and Gemini can be coaxed into providing information that would help with the planning of a terror attack. The GAO points to news reports indicating generative AI helped enable the creation of an explosive device in a January 2025 attack on a Las Vegas hotel and the planning of a February 2026 mass shooting in British Columbia. Lin's team evaluated 11 categories of harmful requests, including child abuse, illegal activities, and fraud, and found uncensored models were generally more likely to comply with harmful requests and generate harmful content than non-uncensored models. He cautioned that any single model's ability to generate harmful content varies with its training data, development methods, and how far its refusal direction was ablated. The hard part for any regulator is intent. When is a question about terrorism a plan, and when is it guidance for writing a thriller? When is a question about hacking intended for theft, and when is it cybersecurity education?

There is a legitimate side. Commercial models are sometimes insufficiently permissive for real security work. Hugging Face, for instance, had to resort to using Z.ai's less rule-encumbered GLM-5.2 to analyze the hacking of its own infrastructure by OpenAI's irresponsibly unsupervised AI agents. Asked whether the safety concerns around uncensored models might prompt a replay of the 1990s, when authorities tried to limit the availability of cryptographic code, Lin said it is difficult to predict whether model abliteration tools will face similar legal challenges. He noted that techniques for both removing and mitigating LLM safety alignment have become active research topics in academia, with many researchers releasing open source tools and artifacts to support their studies.

Lin expects that greater transparency, auditability, and verification of models will help. The current approach taken by commercial labs, limiting powerful models to favored paying insiders and offering hobbled models to the rest, does not work for those left out, and applying a similarly censorious regime to open weight models is unlikely to achieve more success than law enforcement efforts to make cryptography available only to the good guys. The study covered text output only; the issues raised by text-to-image models, including the ability to generate illegal images, were beyond its scope. For executives, the takeaway is that model provenance is now a procurement question. If your product routes to a model you did not train, you may be shipping someone else's refusal settings, someone else's malware signature, and someone else's liability.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Technology