Over 22,000 Uncensored LLMs Circulate on Major Open-Source Platforms Amid Safety and Supply-Chain Concerns
A new academic study reveals that more than 22,000 uncensored large language models hosted across five major platforms lack adequate safety safeguards, with som
An academic study examining major open-source artificial intelligence model sharing platforms has found that more than 22,000 'uncensored large language models (ULLMs)'—whose content-filtering safeguards have been deliberately removed or weakened—are actively circulating online. The findings underscore how the open AI ecosystem faces security and governance gaps similar to those found in traditional open-source software supply chains.

Image source: https://www.theregister.com/ai-and-ml/2026/11/19/uncensored-ai-models-are-thriving-on-code-hosting-platforms/5310891
Researchers Zilong Lin, Zichuan Li, Xiaojing Liao, and XiaoFeng Wang analyzed publicly available model metadata across five prominent model-hosting platforms to assess the prevalence of models lacking safety filters. In this context, uncensored models are those configured to respond to potentially harmful prompts, such as instructions for illegal activities or hacking, without the refusal mechanisms built into commercial offerings.
Over 22,000 Uncensored Models and a Lack of Moderation Warnings
The researchers catalogued 22,096 uncensored large language models across Hugging Face, Kaggle, John Snow Labs, OpenCSG, and ModelScope. Out of all the identified models, only about 13% carried any moderation warning or risk-assessment metadata.
- Massive Distribution Scale: One of the most downloaded uncensored models, 'dolphin-2.9.1-yi-1.5-34b', amassed over 53 million total downloads, with an additional 4.8 million downloads recorded in just the preceding month, demonstrating substantial real-world usage.
- Security Scan Findings: Automated security scans flagged 27 models (accounting for 136,000 total downloads) with threat indicators, such as malware signatures or unsafe deserialization patterns.
Unlike commercial AI services governed by strict safety guardrails, open models can be prompted to produce unrestricted outputs, raising legitimate safety concerns. As previous government reports—such as those by the U.S. Government Accountability Office (GAO)—have highlighted, mainstream models being coaxed into assisting with harmful tasks is an issue extending beyond closed systems into open-weight alternatives.
Cross-Platform Redistribution and Supply-Chain Risks
A central insight of the research is the vulnerability inherent in the open-source AI supply chain. When developers copy and redistribute models across different hosting platforms, safety documentation and risk-assessment metadata are frequently lost in transit.
Without proper moderation warnings or provenance tracking, downstream users and enterprises have little visibility into a model's underlying risk profile. This mirrors well-documented challenges in open-source code repositories where unverified packages are easily imported. Furthermore, the study noted that many uncensored models violate the licensing terms of the original base models from which they were fine-tuned.
Emerging regulatory frameworks, such as the European Union's Artificial Intelligence Act and the Colorado AI Act, increasingly emphasize model documentation and transparency. However, the study provides empirical evidence that current platform practices fall short of these rising regulatory benchmarks.
Responsible Disclosure and Industry Implications
The research team responsibly disclosed their findings to the platforms evaluated in the study. OpenCSG responded constructively, indicating a willingness to improve its services based on the insights. Other platforms, including Hugging Face (reportedly in acquisition talks with Nvidia), Kaggle, and John Snow Labs, did not publicly comment.
Industry analysts suggest this research exposes the friction between open accessibility and safety governance. Unless hosting platforms implement stricter moderation policies, automated malware scanning for model files, and standardized safety labeling, regulatory pressure and security risks surrounding open-weight models are likely to intensify.
Sources
- The Register: Uncensored AI models are thriving on code hosting platforms
- AI Understanding: Study finds thousands of large language models circulating on popular hosting platforms
- U.S. Government Accountability Office (GAO) reports and AI regulatory framework documentation