Uncensored large language models (ULLMs) that can generate harmful content are easily accessible on major model-hosting platforms, according to research from the University of Illinois Urbana-Champaign and Nanyang Technological University. Their study analysed 22,096 ULLMs from five platforms, including Hugging Face, Kaggle, and ModelScope.
These ULLMs are defined as models with absent, removed, or weakened content filtering, allowing them to respond to requests that 'aligned' models would typically refuse. The researchers, including lead author Zilong Lin, found that these models are deployed through applications and GitHub projects, where users may be unaware of the underlying LLM's nature.
Assessing the safety alignment of LLMs is challenging due to the difficulty in examining their behaviour directly from model files. Evaluating safety safeguards typically requires deploying and testing models, which incurs significant computational and financial costs.
The number of ULLMs has risen sharply, from around 100 per month in May 2023 to over 1,000 per month currently. Some of these models have seen millions of downloads, with one model, dolphin-2.9.1-yi-1.5-34b, recording over 53 million downloads.
The study also noted that over 11,000 ULLMs appeared across multiple platforms, with Hugging Face identified as a dominant source. Only about 13 percent of the identified ULLMs carried moderation warnings, and security threat indicators were not consistently preserved across platforms, creating a supply chain problem where risk information can be lost during model copying.
A small number of ULLMs, 27 models downloaded 136,000 times, exhibited threat indicators from security scans, including signatures associated with malware or unsafe deserialization mechanisms.