Researchers Identify 22,096 Uncensored AIs

Researchers have analyzed 22,096 uncensored language models available on five hosting platforms. Their observations focus on the circulation of warnings, generated content, and the security of distributed files. These models also find legitimate uses in security analysis.
Copies Circulate Without All Their Alerts
More than 11,000 models are present on multiple services, says Zilong Lin, lead researcher and assistant professor of computer science at the University of Missouri-Kansas City. Hugging Face is the primary source upstream. About 13% of the listed models came with moderation warnings, but threat indicators do not systematically follow their copies. According to the researchers, dissemination can thus continue without the elements needed to assess risks.
The alerts also concern the files themselves: researchers report threat indicators in 27 models that collectively have 136,000 downloads. Many would contain signatures related to malware, suspicious binary structures, or insecure deserialization mechanisms that could allow arbitrary code execution. Uncensored models also power applications, GitHub projects, and routing services such as Open Router. Their users may be unaware of what type of model meets their requests.
Downloads Reach Several Million
According to the researchers, the monthly rate has increased from about 100 models in May 2023 to over 1,000 currently. Some record millions of downloads. The model dolphin-2.9.1-yi-1.5-34b has surpassed 53 million, with over 4.8 million in the last month alone.
Zilong Lin, Zichuan Li, Xiaojing Liao, and XiaoFeng Wang studied access to models capable of generating harmful content. The team, affiliated with the University of Illinois at Urbana-Champaign and Nanyang Technological University, compiled its corpus from U.S. services Hugging Face, Kaggle, and John Snow Labs, as well as Chinese platforms OpenCSG and ModelScope.
Modified Models to Reduce Refusals
The term ULLM generally covers models whose filters or refusal mechanisms are absent, removed, or weakened. Modest technical skills, or the help of a compliant model, may be enough to modify open-weight models. Toxic training or fine-tuning, referred to as TTFT, and misaligned fine-tuning, known as DTFT, use unsecured data to reduce these refusals; toxicqa-Llama2-13B and WizardLM-7B-Uncensored are among the cited examples. More recently, refusal direction ablation, or ablitération, continues this reduction, as seen in Daredevil-8B-abliterated.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
The researchers tested 11 categories of requests, including those related to child abuse, illegal activities, or fraud. They specify that each model's ability to produce harmful content varies with training data, development methods, and the extent of ablation. Only text responses were evaluated: images, including child exploitation content, were excluded.
However, reduced restrictions may meet a legitimate need when commercial tools refuse authorized security analysis. Hugging Face, for instance, used GLM-5.2 from Z.ai to examine the hacking of its infrastructure by unsupervised OpenAI agents. Conversely, these models can advise on illegal activities, contribute to the creation of malware, or generate manipulative content.
Commercial Tools Also Present Risks
Commercial models presented as aligned are not free from harmful responses, and the criteria for this alignment are not universally agreed upon. The U.S. Government Accountability Office, or GAO, indicates that Claude, Grok, ChatGPT, and Gemini may provide useful information for preparing an attack.
The GAO also relays reports attributing generative AI's contribution to the creation of an explosive device during an attack on a Las Vegas hotel in January 2025. It also cites reports regarding its assistance in preparing a mass shooting in British Columbia, Canada, in February 2026.
Costly Checks and OpenCSG's Response
For Zilong Lin, model hosts serve a function similar to GitHub, GitLab, or Codeberg: enabling sharing, modification, and redistribution. However, one difference complicates checks: unlike software behavior, which is often examinable in its code, the security of a model is difficult to deduce from its files. Generally, it must be deployed and then subjected to potentially harmful queries. According to the researcher, the computational and financial cost of these operations complicates their execution across large catalogs.
The researchers also report numerous violations of the licenses of original models. Lin claims to have submitted the findings through a responsible disclosure process to Hugging Face, John Snow Labs, OpenCSG, Kaggle, ModelScope, and Open Router.
Lin reports that OpenCSG praised the work of identification and measurement, as well as its interest in a safer and more reliable ecosystem. OpenCSG has committed to improving its service based on these results.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.