Policy

22,000 Uncensored AI Models Found on Major Hosting Platforms

Academic study reveals governance gaps as unfiltered language models spread across Hugging Face, Kaggle, and other repositories.

Omega Editorial · November 19, 2026 · 3 min read

Key takeaways

  • Researchers identified 22,096 uncensored AI models across five major hosting platforms, with only 13% carrying moderation warnings.
  • Monthly production of uncensored models has grown from 100 in May 2023 to over 1,000 today, with some models downloaded tens of millions of times.
  • Twenty-seven models contained security threats including malware signatures and mechanisms for arbitrary code execution, downloaded 136,000 times combined.
  • Over 11,000 models appeared on multiple platforms with Hugging Face as the primary source, creating a supply chain problem where risk information gets lost during transfers.
  • The research highlights a governance gap where models travel easily between platforms but safety documentation does not follow them.

More than 22,000 uncensored large language models capable of producing harmful content are readily available on popular code hosting platforms, according to new research that exposes significant gaps in AI governance.

Researchers from the University of Illinois Urbana-Champaign and Nanyang Technological University analyzed models from five major hosting platforms: Hugging Face, Kaggle, John Snow Labs, OpenCSG, and ModelScope. Their findings, detailed in a paper titled "Characterizing the Ecosystem of Open-Source Uncensored Large Language Models: Supply Chain, Governance, and Risks," reveal an ecosystem where safety information fails to travel with the models themselves.

Why it matters

The proliferation of uncensored models creates a supply chain problem for AI safety. When models move between platforms, critical risk information gets lost. Only 13 percent of identified models carried moderation warnings, meaning users deploying these models through applications or GitHub projects may be unaware of their unfiltered nature. This governance gap undermines industry safety commitments and complicates efforts to prevent misuse.

Explosive growth in unfiltered models

The volume of uncensored LLMs has surged from roughly 100 per month in May 2023 to more than 1,000 monthly today. Some models have achieved massive reach—the uncensored model dolphin-2.9.1-yi-1.5-34b has logged over 53 million downloads, including 4.8 million in the past month alone.

Uncensored models are created through techniques like toxic training/fine-tuning, de-aligned training, and refusal-direction ablation (nicknamed "abliteration"). These methods remove or weaken the safety guardrails that prevent aligned models from responding to harmful prompts.

"Models can travel more easily than the information needed to assess their risks," said Zilong Lin, lead author and assistant professor at the University of Missouri-Kansas City, according to The Register. Over 11,000 models appeared across multiple platforms, with Hugging Face serving as the dominant upstream source.

Security threats and licensing violations

Twenty-seven models exhibited threat indicators from security scans, including malware signatures, suspicious binary patterns, or unsafe deserialization mechanisms that could enable arbitrary code execution. These 27 models have been downloaded 136,000 times combined.

Many uncensored models also violate the licensing terms of their base models, though such violations have historically had little impact on AI industry practices.

Legitimate uses complicate regulation

The challenge with restricting uncensored models lies in determining user intent. Security researchers sometimes need less restrictive models for legitimate analysis—Hugging Face itself recently used a more permissive model to analyze a security breach because commercial models refused the queries.

Lin expects greater transparency and verification standards to help address risks. The EU AI Act and Colorado AI Act both emphasize documentation requirements, including model cards that detail capabilities, limitations, and evaluation results.

The researchers disclosed their findings to all platforms studied. OpenCSG responded positively and committed to enhancing its service based on the research. Hugging Face, Kaggle, and John Snow Labs did not respond to requests for comment.

The findings were first reported by The Register.

  • #ai safety
  • #uncensored llms
  • #hugging face
  • #model governance
  • #open source ai
  • #ai security

This is an original analysis by the Omega editorial team. Source reporting: AI Watch.

Want systems like this working for your business?

Book a Call

More in Policy