Uncensored AI Models: 22,000 Found on Hosting Sites

Uncensored AI models spreading across model-hosting platforms, shown as a supply chain

Uncensored AI models that can produce harmful content are easy to find on the big model-hosting sites. Researchers at the University of Illinois Urbana-Champaign and Nanyang Technological University in Singapore analyzed 22,096 candidates from five platforms and identified more than 11,000 of them as uncensored.

The research paper on this open-source ecosystem, covering its supply chain, governance and risks, has four authors: Zilong Lin, Zichuan Li, Xiaojing Liao and XiaoFeng Wang.

Where the models live

The five platforms are Hugging Face, Kaggle and John Snow Labs in the US, and OpenCSG and ModelScope in China. Hugging Face is the dominant upstream source. The models then spread into apps, GitHub projects and routing services such as Open Router, so people using those services may not know which model sits underneath.

The authors treat a model as uncensored when its filters or refusal safeguards are missing, stripped out or diluted. Lin, the lead author and an assistant professor at the University of Missouri-Kansas City, compares model hosts to code hosts like GitHub. Both enable sharing, modification and redistribution. The difference is that a model’s safety behavior can’t be read from its files. Checking it means running the model and testing harmful prompts, which costs money and compute.

How the models get made

Anyone with modest technical skill can strip refusals from an open-weight model. Fine-tuning on unsafe data, whether toxic or de-aligned, yielded models like toxicqa-Llama2-13B and WizardLM-7B-Uncensored. A newer technique called refusal-direction ablation, or “abliteration,” weakens the refusal training directly, and Daredevil-8B-abliterated is one example. The technique appeared in 2024, and public tools for it have multiplied in about two years.

The scale and the warnings gap

Growth is steep. Monthly uploads went from around 100 in May 2023 to above 1,000 today. Several have been downloaded millions of times. The model dolphin-2.9.1-yi-1.5-34b alone has passed 53 million downloads, with over 4.8 million in the past month.

Roughly 13% of the models had a moderation warning, and security threat flags often didn’t survive the move from one platform to another. Lin calls this a supply chain problem: a copy can travel while the information needed to judge its risks gets lost.

A small group, 27 models downloaded 136,000 times, showed threat indicators in security scans. Many had malware signatures, suspicious binary patterns or unsafe deserialization that could allow arbitrary code execution. Many of the models also violate the license terms of the models they derive from.

How risky they are

The team tested text responses to harmful requests in 11 categories, among them child abuse, illegal activity and fraud. The uncensored models were generally more likely to comply than non-uncensored ones. Results varied with training data, method and how far the refusal direction was removed. Image generation was outside the study.

Commercial models are not immune either. A US Government Accountability Office report found that Claude, Grok, ChatGPT and Gemini can each be talked into providing help with planning a terror attack.

Legitimate uses and hard tradeoffs

Less restrictive models have real uses. Security analysis is one: commercial models are sometimes so cautious that they refuse legitimate work. Telling a plot for a thriller from a plan, or hacking education from theft, depends on guessing intent. And restricting open weights the way governments once tried to restrict cryptography may not work any better.

What researchers want next

Lin points to transparency, auditability and verification. The team found these models through public metadata such as model cards, which shows documentation is useful. He cites the EU AI Act, which requires technical documentation for general-purpose models, and the Colorado AI Act, which recognizes model cards in impact assessments.

Before publishing, the team notified Hugging Face, John Snow Labs, OpenCSG, Kaggle, ModelScope and Open Router about its findings before publication. Only OpenCSG replied, saying the research is meaningful and committing to improve its service. Hugging Face, Kaggle and John Snow Labs did not respond to requests for comment. Whether the host sites add stronger labeling is now the open question for uncensored AI models.

Source: theregister

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.