Back to News
SecurityAI Understanding briefing

Study finds thousands of uncensored large language models circulating on popular hosting platforms

Researchers from the University of Illinois Urbana‑Champaign and Nanyang Technological University identified over 22,000 uncensored LLMs on sites like Hugging Face and Kaggle, highlighting a governance gap and safety risks.

5 min readRead the original reporting
Source-provided image accompanying Study finds thousands of uncensored large language models circulating on popular hosting platforms
Attributed reportingSource recorded
Publisher
theregister.com
Source type
Reporting by a news outlet — not a first-party document.

What we could not confirm independently: This claim is attributed to the named outlet. We did not verify it against a first-party document. (theregister.com)

ContextUnderstand this in 60 seconds

Key terms

Weight
A learned numeric value that scales signals passing through a neural network.
AI Act
The European Union's risk-based regulatory framework for AI systems and providers.
Test yourselfAI Ethics Quiz

What happened

A new academic paper catalogued more than 22,000 uncensored large language models (ULLMs) hosted on five major model‑hosting platforms, revealing that many lack safety safeguards and can be used to generate harmful content. The study, authored by Zilong Lin, Zichuan Li, Xiaojing Liao and XiaoFeng Wang, examined publicly available model metadata to identify models that either omit or deliberately weaken content‑filtering mechanisms. Researchers found that roughly 13 % of the identified models carried any moderation warning, and a small subset showed security‑related threat indicators such as malware signatures. The most downloaded uncensored model, dolphin‑2.9.1‑yi‑1.5‑34b, had amassed over 53 million downloads, with 4.8 million in the last month alone. The authors warned that the ease of copying models across platforms erodes the traceability of safety documentation, creating a supply‑chain risk similar to that seen in open‑source software.

The researchers collected 22,096 uncensored LLMs from five major hosting services—Hugging Face, Kaggle, John Snow Labs, OpenCSG and ModelScope—by analyzing publicly available model cards and metadata. They defined uncensored models as those whose content‑filtering safeguards are absent, removed, or weakened, allowing them to respond to harmful prompts.

Only about 13 % of the identified models included any moderation warnings, and security scans flagged 27 models (136,000 total downloads) with threat indicators such as malware signatures or unsafe deserialization patterns. The most popular uncensored model, dolphin‑2.9.1‑yi‑1.5‑34b, recorded more than 53 million downloads overall, with a recent surge of 4.8 million downloads in a single month.

The study evaluated the models on 11 categories of harmful requests (including child abuse, illegal activity, and fraud) and found that uncensored models were generally more likely to comply with such requests than aligned counterparts. However, compliance rates varied based on training data, fine‑tuning techniques, and the extent of refusal‑direction ablation applied.

The authors highlighted a supply‑chain issue: models can be copied across platforms without preserving the accompanying risk‑assessment metadata, making it difficult for downstream users to gauge safety. They also noted that many uncensored models violate the licensing terms of the base models they were derived from.

The research team responsibly disclosed their findings to the platforms studied. OpenCSG responded positively, indicating a willingness to improve its service based on the results. Other platforms, including Hugging Face (currently in acquisition talks with Nvidia), Kaggle, and John Snow Labs, did not comment.

Source details: theregister.com

Why it matters

The proliferation of uncensored LLMs poses a concrete safety and security threat because these models can comply with requests that aligned commercial models would refuse, including instructions for illegal activities, weaponization, or fraud. The study’s findings intersect with recent government reports—such as the U.S. GAO’s identification of mainstream models being coaxed into providing terror‑planning advice—demonstrating that the problem is not limited to open‑source models. The lack of consistent moderation warnings (only about 13 % of models) and the loss of risk‑assessment metadata when models are copied across platforms create a governance gap that hampers regulators, enterprises, and end‑users from assessing model risk. Moreover, the discovery that a handful of models exhibit malware‑like signatures raises the specter of code‑execution attacks via model loading. The research underscores the need for stronger documentation standards, such as those mandated by the EU and Colorado AI Act, and suggests that transparency mechanisms like model cards are essential for tracking safety attributes across the model supply chain.

Safety risk: Uncensored models can be weaponized to produce instructions for terrorism, hacking, or fraud, as demonstrated by prior GAO reports linking mainstream models to real‑world attacks. The study shows that open‑ models present a similar, if not greater, risk due to the lack of built‑in refusal mechanisms.

Governance gap: The loss of moderation warnings and risk metadata when models are redistributed creates a blind spot for regulators and enterprises that rely on documentation to assess AI risk. This mirrors challenges seen in other open‑source ecosystems where code provenance is hard to trace.

Security threat: The detection of malware‑like signatures in a subset of models suggests a potential attack vector where loading a model could lead to arbitrary code execution, expanding the threat surface beyond content generation.

Regulatory relevance: The findings align with emerging AI legislation that emphasizes model documentation (EU , Colorado AI Act). The study provides empirical evidence that current documentation practices are insufficient for open‑ models.

Industry impact: Hosting platforms may need to revise their policies, implement stricter moderation, or develop automated safety‑assessment tools to maintain user trust and comply with forthcoming regulations.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

System Requirements:
Best ArchitecturePure RAGRecommended pattern
Hallucination RiskVery LowGrounding efficacy
Update Cost$0 (Vector sync)Ongoing maintenance
Core takeaway: Fine-tuning teaches models how to speak (form, style, syntax); RAG teaches models what to say (verifiable facts). Never use fine-tuning alone for factual memory.
Interactive Concept Check+10 Points
AI Ethics Quiz

Why can ethical evaluation not be reduced to one model score?

What to watch next

Future regulatory actions targeting model‑card transparency, industry responses from hosting platforms, and the development of tools to detect and flag uncensored models are key signals to monitor. Watch for any policy proposals that extend the EU ’s documentation requirements to open‑ models, as well as any technical solutions from academia or industry aimed at automating safety‑assessment of model files. Additionally, observe whether major platforms such as Hugging Face or Kaggle introduce stricter moderation or labeling policies in response to the study’s disclosures.

Policy developments: Legislative bodies may extend documentation and safety‑assessment requirements to open‑ models, potentially mandating standardized model cards or safety certifications.

Platform responses: Hugging Face, Kaggle, and other hosting services could introduce mandatory safety labeling, automated scanning for harmful content, or restrictions on model uploads that lack moderation warnings.

Technical solutions: Academic and industry groups may release open‑source tools for automated detection of uncensored behavior or malware signatures in model files, influencing best‑practice standards.

Legal actions: Similar to historical attempts to restrict cryptographic software, there could be legal challenges aimed at tools that facilitate the removal of safety safeguards from LLMs.

Adoption trends: Monitoring download statistics for identified uncensored models will indicate whether the community shifts toward safer alternatives or continues to favor unrestricted models for niche use cases.

Related guides & quizzes

Found this useful?