Uncensored large language models (ULLMs) capable of producing harmful content are readily available at popular model-hosting websites, calling into question AI industry commitments to safety.
Researchers affiliated with University of Illinois Urbana-Champaign in the US and Nanyang Technological University in Singapore have analysed 22,096 ULLMs collected from five major LLM hosting platforms.
As described in their paper, "Characterizing the Ecosystem of Open-Source Uncensored Large Language Models: Supply Chain, Governance, and Risks," authors Zilong Lin, Zichuan Li, Xiaojing Liao, and XiaoFeng Wang found that uncensored LLMs are easy to obtain from hosting platforms like US-based Hugging Face, Kaggle, and John Snow Labs, and China-based OpenCSG and ModelScope.
What's more, these unrestrained models have been deployed through applications, GitHub projects, and model routing platforms like Open Router, and users of those apps or services may not be aware of the nature of the underlying LLM.
"Uncensored LLMs generally refer to LLMs whose content filtering or refusal safeguards are absent, removed, or weakened, allowing them to respond to requests that aligned models may refuse," the paper explains.
Zilong Lin, lead author and assistant professor of computer science at the University of Missouri-Kansas City, told The Register that open source LLM hosting platforms like Hugging Face play a similar role to code hosting services like GitHub, GitLab, and Codeberg in the open source community by enabling sharing, modification, and redistribution of code artifacts.
"However, unlike traditional software, whose behavior can often be examined through source code analysis, the safety alignment of LLMs is difficult to assess directly from their model files," Lin explained. "Evaluating whether an LLM has adequate safety safeguards typically requires deploying the model and testing its responses to potentially harmful prompts, which involves substantial computational and financial costs."
Lin said this makes the identification and assessment of LLM safety particularly challenging, particularly given the large number of models available. That's why he and his colleagues chose to study the space through publicly available model metadata.
Even popular commercial LLMs that are purportedly "aligned" – to some standard not everyone accepts – may respond to prompts with content deemed harmful. The US Government Accountability Office just released a report that found Claude, Grok, ChatGPT, and Gemini can be coaxed to provide information that would help with the planning of a terror attack.
To highlight the consequences of AI advice, the GAO points to news reports indicating that 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, Canada.
The issue with open weight LLMs is that anyone with modest technical skills (or just cluelessly assisted by a compliant LLM) can apply techniques like toxic training/fine-tuning (TTFT) or de-aligned training/fine-tuning (DTFT) to datasets that haven't been safe, in order to produce models with reduced or removed refusal behavior. Examples include toxicqa-Llama2-13B and WizardLM-7B-Uncensored. A more recent technique called refusal-direction ablation (or abliteration, e.g., Daredevil-8B-abliterated) dilutes refusal training to make the model less likely to refuse requests.
Uncensored (or simply less restrictive) models have legitimate uses, sometimes ironically because commercial models are insufficiently permissive and refuse legitimate security analysis. Hugging Face, for example, had to resort to using Z.ai's less rule-encumbered GLM-5.2 to analyze the hacking of its infrastructure by OpenAI's irresponsibly unsupervised AI agents.
But ULLMs can just as easily be misused to provide guidance about illegal activity, assist with malware development, or emit content that's harmful, malicious, or manipulative in some way.
"We identified more than 11,000 ULLMs appearing across multiple platforms, with Hugging Face serving as the dominant upstream source," Lin said in a blog post.
"This movement exposes an important governance gap: models can travel more easily than the information needed to assess their risks. Only about 13 percent of the ULLMs we identified carried moderation warnings, and security threat indicators were not consistently preserved across platforms.
"This creates a supply chain problem: a model can be copied across platforms while the information needed to understand its risks is lost along the way."
The ULLMs were evaluated for text output and not images. The issues raised by text-to-image LLMs, such as the ability to generate illegal images (e.g. CSAM), were beyond the scope of this particular study.
"The uncensored LLMs we evaluated were generally more likely to comply with harmful requests and generate harmful content than non-uncensored models," Lin said. "Our evaluation covered 11 categories of harmful requests, including child abuse, illegal activities, and fraud.
"However, the ability of an individual model to generate harmful content varies depending on several factors, including its training data, development methods, and the extent to which its refusal direction has been ablated."
A small portion of these ULLMs (27 of them, which have been downloaded 136,000 times) exhibited threat indicators from security scans. Many of these, authors say, had signatures associated with malware, exhibited suspicious binary patterns, or contained unsafe deserialization mechanisms that could allow arbitrary code execution.
The number of ULLMs has been rising rapidly, according to the paper, from about 100 per month in May 2023 to more than 1,000 per month presently. And some of these models have been downloaded millions of times. The uncensored model dolphin-2.9.1-yi-1.5-34b has more than 53 million downloads, and more than 4.8 million just last month.
The authors note that many ULLMs violate the licensing terms of the models on which they've been based, not that source code licensing has had any meaningful impact on the behavior of the AI industry.
Asked whether the safety concerns arising from the availability of ULLMs 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.
"As shown in Section 5.1 of our paper, refusal-direction ablation (i.e., abliteration) appeared in 2024 and quickly gained popularity within the open-source LLM community. Within approximately two years, a growing ecosystem of publicly available tools had emerged to support this technique.
"In the meantime, techniques for both removing and mitigating LLM safety alignment have also become active research topics in academia, with many researchers releasing open-source tools and artifacts to support their studies."
Dealing with the risks posed by uncensored models isn't easy because it depends on divining the intent of the user's prompt. When is a question about terrorism intended as 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?
The current approach taken by commercial labs — limiting powerful models to favored paying insiders and offering hobbled models to the rest — doesn't work for those left out. Applying some similarly censorious regime to open weight models isn't likely to achieve any more success than efforts by law enforcement to make cryptography available only to "the good guys."
Lin expects that greater transparency, auditability, and verification of models will help.
"As discussed in Section 8 of our paper, our approach to identifying uncensored LLMs relies on publicly available model metadata, such as model cards, to identify models that explicitly disclose their uncensored characteristics," he said. "Our findings demonstrate the value of such transparency for identifying potentially unsafe models.
"Regulatory frameworks have also increasingly emphasized AI documentation and transparency. For example, the EU AI Act requires providers of general-purpose AI models to maintain technical documentation covering model characteristics, capabilities, limitations, and evaluation results. Similarly, the Colorado AI Act recognizes model cards and dataset cards as possible documentation supporting AI impact assessments."
Hugging Face, in the process of being acquired by Nvidia, did not respond to a request for comment. Nor did Kaggle or John Snow Labs, the two other US-based companies named in the study.
According to Lin, the research findings were responsibly disclosed to Hugging Face, John Snow Labs, OpenCSG, Kaggle, ModelScope, and Open Router.
Asked whether any of the companies responded, Lin said, "We received a response from OpenCSG, which expressed appreciation for our work on identifying and measuring uncensored LLMs, stating that 'your research is meaningful and supports the development of a safe and trustworthy LLM ecosystem.'"
OpenCSG, the report says, has committed to enhance its service based on the research findings. ®