Researchers find warnings on just 13% of sampled uncensored AI models
The study separates two risks: models more willing to answer harmful requests, and a small subset with security threat indicators.
Loading page…
The study separates two risks: models more willing to answer harmful requests, and a small subset with security threat indicators.
Listen to this story
Researchers led by Zilong Lin traced a weakening-safeguard ecosystem beyond model repositories, into apps, public GitHub projects and routing services, where copies may lose threat indicators. The study identifies two routes for reducing refusals: filtering alignment data and using refusal-direction ablation, now supported by public automation. Separate scans found security indicators in 27 models with 136,000 combined downloads; those flags do not establish that every model contained malware. OpenCSG pledged improvements, which had not yet been completed.
Only about 13% of identified models carried moderation warnings, while about 70% advertised uncensored characteristics.
Publishers claimed more than 11,000 models could perform harmful functions, but those descriptions do not prove capability.
Researchers found 167 datasets that removed alignment-related content and 283 containing harmful or toxic material.
Uncensored AI models are easy to find; their risk information is less reliable. Researchers identified 22,096 models across five hosting platforms, but only about 13% carried moderation warnings. In findings described by lead author Zilong Lin on November 17, models also moved between platforms without consistently retaining security threat indicators.
The sample covers models identified as of March 1, 2026—not a live November count. More than 2,100 developers created them across Hugging Face, OpenCSG, ModelScope, Kaggle and John Snow Labs. The researchers identified uncensored models through public descriptions and other metadata, rather than treating every downloadable model as uncensored.
“Uncensored” describes models whose refusal safeguards or content filters are absent, weakened or removed. The label does not mean every model responds the same way. About 70% of the identified models advertised uncensored characteristics in public metadata, making their positioning visible to anyone browsing their descriptions.
More than 11,000 claimed functions associated with harmful content, including malicious-code generation and offensive cybersecurity. Those descriptions are claims by model publishers, not proof that each model can perform the advertised tasks. Separately, the researchers found more than 11,000 models appearing on multiple platforms, with Hugging Face the dominant upstream source.
The researchers found uncensored characteristics promoted in public metadata for about 70% of identified models.
Only about 13% of the identified models carried moderation warnings.
Distribution extends beyond model repositories. The researchers found uncensored models powering web applications, public GitHub projects, services discussed on underground forums and model-routing platforms. These routing services offer access to multiple models through one software connection, or API. Lin’s account describes a supply-chain gap: model copies can travel while information needed to assess their risks gets lost.
The study also traced the materials used to develop these models. It identified 970 referenced datasets, including two groups that show how developers can change a model’s behavior:
Another method, refusal-direction ablation, weakens refusal behavior without conventional fine-tuning—the process of further training a model. Lin says publicly available tools now automate this method, reducing the expertise and resources needed. That provides a different route from changing training datasets: developers can instead alter the model’s refusal behavior more directly.
In response testing, uncensored models were generally more likely than non-uncensored models to comply with harmful requests and generate harmful content, Lin told The Register. The evaluation covered 11 categories, including child abuse, illegal activities and fraud. It assessed text output, not image generation; individual results varied with training data, development methods and the extent of refusal removal.
Security scans flagged a separate issue in 27 models with a combined 136,000 downloads. Indicators included malware-associated signatures, suspicious binary patterns and unsafe deserialization mechanisms—ways of loading stored data that could allow arbitrary code execution. These were security threat indicators, not a finding that all uncensored models contained malware.
Lin said the team disclosed its findings to all five hosting platforms and OpenRouter. OpenCSG responded and committed to improving its service based on the research. That is a commitment to act, rather than a completed fix.
Loading discussion...
Join the conversation
Explain what a separate warning would—or wouldn’t—add.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.