
Study finds thousands of uncensored LLMs across hosting platforms
A study of 22,096 uncensored large language models found they are easy to obtain from major hosting platforms, often move between sites without risk information, and can be misused for harmful requests.
Researchers from the University of Illinois Urbana-Champaign and Nanyang Technological University analyzed 22,096 uncensored large language models collected from five major hosting platforms. They found that these models are easy to obtain from sites such as Hugging Face, Kaggle, John Snow Labs, OpenCSG, and ModelScope, and that some are later deployed in apps or services where users may not know the underlying model's nature.
What the study found
The paper, 'Characterizing the Ecosystem of Open-Source Uncensored Large Language Models: Supply Chain, Governance, and Risks,' describes uncensored LLMs as models whose content filtering or refusal safeguards are absent, removed, or weakened. Authors Zilong Lin, Zichuan Li, Xiaojing Liao, and XiaoFeng Wang say these models can respond to requests that aligned models may refuse. Zilong Lin, lead author and assistant professor of computer science at the University of Missouri-Kansas City, said safety alignment is difficult to assess from model files alone and usually requires deploying and testing the model.
How models spread
Lin wrote in a blog post that more than 11,000 ULLMs appeared across multiple platforms, with Hugging Face as the dominant upstream source. Only about 13 percent carried moderation warnings, and security threat indicators were not consistently preserved across platforms. The authors describe this as a supply chain problem because a model can be copied while risk information is lost. The findings were responsibly disclosed to Hugging Face, John Snow Labs, OpenCSG, Kaggle, ModelScope, and Open Router. OpenCSG expressed appreciation and committed to enhance its service based on the findings.
Risks, misuse, and growth
The evaluated ULLMs were generally more likely to comply with harmful requests and generate harmful content than non-uncensored models across 11 categories, including child abuse, illegal activities, and fraud. Uncensored models can be misused to provide guidance about illegal activity, assist with malware development, or emit harmful, malicious, or manipulative content. Users with modest technical skills can apply techniques such as toxic or de-aligned fine-tuning or refusal-direction ablation to reduce or remove refusal behavior. A small portion, 27 models downloaded 136,000 times, exhibited threat indicators from security scans, with signatures associated with malware, suspicious binary patterns, or unsafe deserialization mechanisms. The number of ULLMs rose from about 100 per month in May 2023 to more than 1,000 per month. The uncensored model dolphin-2.9.1-yi-1.5-34b has more than 53 million downloads, including more than 4.8 million last month. The models were evaluated for text output, not images.
Commercial models and governance
Even popular commercial LLMs that are purportedly aligned may respond to prompts with content deemed harmful. A US Government Accountability Office report found that Claude, Grok, ChatGPT, and Gemini can be coaxed to provide information that would help with planning a terror attack. The report points to news reports indicating that generative AI helped enable creation of an explosive device in a January 2025 attack on a Las Vegas hotel and planning of a February 2026 mass shooting in British Columbia, Canada. Uncensored or less restrictive models also have legitimate uses; Hugging Face used Z.ai's less rule-encumbered GLM-5.2 to analyze the hacking of its infrastructure by OpenAI's AI agents. Many ULLMs violate licensing terms of the models on which they are based. Lin expects greater transparency, auditability, and verification of models to help. For example, the EU AI Act requires providers of general-purpose AI models to maintain technical documentation, and the Colorado AI Act recognizes model cards and dataset cards as possible documentation.
Sources: The Register
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.