Hugging Face Partners with Baseten and Goodfire on Open-Weight AI Safety
Baseten's research arm, Base Labs, has partnered with Hugging Face and Goodfire AI to develop safety evaluation, training, and monitoring infrastructure for open-weight AI models. The initiative explicitly addresses security concerns surrounding open-weight models, including those modified to remove safety filters using techniques like abliteration. Because Hugging Face hosts thousands of abliterated model weights, open-source AI developers fear this partnership could signal a shift toward stricter hosting or moderation policies against uncensored models. However, establishing public safety evaluation standards could also help open-weight models gain broader trust and regulatory acceptance. The announcement specifically highlights that Hugging Face hosts over 6,000 abliterated models, raising questions about how automated safety tools will evaluate uncensored weights. While the effort focuses on developing public evaluation methods, Hugging Face has not officially announced any plans to ban or remove abliterated repositories.
## BACKGROUND
"Abliteration" is a technique that removes a large language model's built-in refusal mechanism by altering internal representation vectors without requiring full retraining. While developers use uncensored models for research, creative writing, and unconstrained assistant tasks, safety researchers view them as high-risk because they bypass built-in safeguards. Hugging Face serves as the primary distribution hub for open-source AI, making any potential moderation changes highly consequential for the ecosystem.