Back to Home

22,096 Uncensored LLMs, and Only 13% Flagged

The open-weight ecosystem has a moderation problem that is now measurable. Researchers at the University of Illinois Urbana-Champaign and Nanyang Technological University in Singapore catalogued 22,096 uncensored large language models across five hosting platforms, then tracked how those models travel between services.

Their paper treats de-aligned weights as a supply chain problem, not a content moderation edge case.

The result is less a scandal than a structural finding. The models are legal, mostly labeled, widely downloaded, and rarely warned about.

Uncensored here has a precise meaning: models whose refusal safeguards are absent, removed, or weakened, so they answer requests that aligned models decline.

Inside the 22,096-model supply chain

The team studied public metadata rather than outputs. Model cards and tags named the behavior explicitly often enough to build a census.

That choice was forced. Alignment cannot be inspected from weights the way a bug can be found in source code, and probing 22,000 models costs real GPU time.

The five platforms split cleanly by geography:

  • Hugging Face, Kaggle, and John Snow Labs, all US-based
  • OpenCSG and ModelScope, both China-based
  • Hugging Face acting as the dominant upstream source for the rest

Lead author Zilong Lin, an assistant professor at the University of Missouri-Kansas City, compared these hosts to GitHub and GitLab: infrastructure for sharing and redistributing artifacts.

The difference is that a fork of a repository can be audited. A fork of a model cannot be audited without running it.

About 11,000 of the catalogued models appear on more than one platform, and the flow runs outward from Hugging Face.

Volume is climbing. The count moved from about 100 new uncensored models per month in May 2023 to more than 1,000 per month today.

Demand is not hypothetical. The model dolphin-2.9.1-yi-1.5-34b has passed 53 million downloads, with 4.8 million in the last month alone.

Production has been industrialized. Three techniques account for most of the supply:

  • Toxic fine-tuning, or TTFT: training on datasets that were never safety-filtered, as in toxicqa-Llama2-13B
  • De-aligned fine-tuning, or DTFT: deliberately training refusal behavior away, as in WizardLM-7B-Uncensored
  • Refusal-direction ablation, or abliteration: editing weights along a computed refusal direction, as in Daredevil-8B-abliterated

Abliteration matters most for capability diffusion. It needs no preference data and no RLHF pipeline, only a direction vector and a weight edit.

Lin noted the technique appeared in 2024 and that a full tooling ecosystem grew around it within about two years.

It also has defenders. Commercial models are restrictive enough that Hugging Face used Z.ai's GLM-5.2 to analyze the hacking of its own infrastructure by OpenAI agents.

Where the warnings get lost

The governance finding is sharper. Only about 13 percent of the catalogued models carried any moderation warning.

Risk signals do not survive the copy. A model mirrored to another host frequently arrives without the flags, notes, or scan results attached to the original.

Scanning found 27 models carrying threat indicators, collectively downloaded about 136,000 times: malware associations, suspicious binary patterns, and unsafe deserialization that allows arbitrary code execution on load.

That last item deserves attention. Older pickle checkpoints execute code during deserialization, which is why safetensors exists. A model card warning is not a sandbox.

Distribution compounds it. These weights reach users through apps, GitHub projects, and routing services like OpenRouter, and those users often have no idea what is answering them.

Behavioral testing covered 11 categories of harmful requests, including child abuse, illegal activity, and fraud. Uncensored models complied more often than aligned ones, with wide variation by model.

Variation tracked training data, the development method, and how thoroughly the refusal direction had been ablated.

Intent is the hard part. A question about terrorism may be a plot or a thriller outline; a question about exploitation may be theft or defensive research.

Closed models share the flaw. A recent GAO report found Claude, Grok, ChatGPT, and Gemini could be coaxed into material useful for planning a terror attack.

The boundary is not open versus closed. It is measured versus assumed.

Current practice, gating strong models behind paid tiers and shipping restricted versions to everyone else, leaves excluded users to find substitutes.

Asked whether abliteration tools might face the legal pressure cryptography met in the 1990s, Lin said it is difficult to predict.

What would actually help

His preferred lever is documentation, not restriction: transparency, auditability, and verification do more than takedowns.

The method depended on operators volunteering the truth in model cards, a weak dependency but a real one.

Regulation already points that way. The EU AI Act requires general-purpose providers to document capabilities, limitations, and evaluation results, and Colorado accepts model cards in impact assessments.

Three fixes follow:

  • Portable risk metadata, so a warning travels with the weights across mirrors and routers
  • Standardized model cards recording the ablation method and the categories tested
  • Artifact-level scanning for unsafe deserialization and binary anomalies, separate from behavioral testing

Licensing enforcement is not on that list. Many of these models violate the terms of the weights they were derived from, and those terms have not changed industry behavior.

Findings were disclosed to all six platforms and the routing service. Only OpenCSG responded, calling the work meaningful and committing to improve its service.

Hugging Face, Kaggle, and John Snow Labs did not comment. Hugging Face is being acquired by Nvidia, which makes moderation there a corporate question as much as a policy one.

The number to watch is not 22,096. It is 13 percent, the share of an ecosystem that documents its own risk.

Comments

No comments yet. Be the first to share your thoughts!