2 Sources
[1]
Uncensored AI models are thriving on code hosting platforms
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
[2]
Study finds 22,096 uncensored AI models on major hosting platforms
Researchers at the University of Illinois Urbana-Champaign and Nanyang Technological University identified 22,096 uncensored large language models across five major model-hosting platforms, finding that many lack clear safety warnings. The models were available through U.S.-based Hugging Face,
Share
Copy Link
Researchers counted 22,096 AI models with their safety safeguards stripped out across five major hosting platforms, and found most of them carry no warning labels at all
1
.
Source: The Register
The team, from the University of Illinois Urbana-Champaign and Nanyang Technological University, defines uncensored models as ones whose content filters or refusal safeguards are missing, stripped out or weakened
2
. That means they may answer things ordinary models refuse2
. The five platforms they examined were Hugging Face, Kaggle and John Snow Labs in the US, and OpenCSG and ModelScope in China2
. These are mainstream services, not fringe corners of the internet.Hugging Face, a major open-source model hub now being acquired by Nvidia, served as the dominant upstream source, with more than 11,000 of the identified models appearing on several platforms at once
2
. Only about 13 percent carried moderation warnings1
.The ecosystem is also growing fast. Monthly output has grown from about 100 new uncensored models in May 2023 to more than 1,000 currently
2
. One model, dolphin-2.9.1-yi-1.5-34b, has been downloaded more than 53 million times, over 4.8 million of those in the past month2
.When the researchers tested text responses across 11 categories of harmful requests, from child abuse to fraud, the uncensored models generally proved likelier to comply than ordinary ones
1
.Harm varies model by model, depending on training data, how the model was built and how much of its alignment was stripped away
1
. The findings cover text output only; image generation was out of scope1
.A separate risk showed up in security scans: 27 models, downloaded a combined 136,000 times, carried threat indicators such as malware-linked signatures or flaws that could allow arbitrary code to run
2
.Related Stories
The team worked from publicly available model metadata. Testing a model's safety the direct way means deploying it and probing it with harmful prompts, which costs serious computing power and money
1
2
. Safeguard removal itself is easy to come by: a technique called abliteration emerged in 2024 and spread quickly through the open-source community2
.The picture is not purely one-sided. The US Government Accountability Office found that Claude, Grok, ChatGPT and Gemini could all be coaxed into providing information useful for planning a terror attack
2
. Less-restricted models have legitimate uses too. Hugging Face used a less-restricted model from Z.ai to analyze a hack of its own infrastructure by OpenAI's AI agents, because commercial models were too tightly filtered for the job1
.The researchers notified all five platforms plus Open Router. Only OpenCSG replied, saying the research "is meaningful and supports the development of a safe and trustworthy LLM ecosystem"
1
. Hugging Face, Kaggle and John Snow Labs did not respond to requests for comment1
."This movement exposes an important governance gap: models can travel more easily than the information needed to assess their risks," Lin said
1
. He suggested transparency requirements, such as the technical documentation the EU AI Act already demands of general-purpose AI models, could help surface unsafe models1
.Summarized by
Navi
[1]
29 Jan 2026•Technology

28 Jul 2026•Technology

26 May 2026•Technology
