Base Labs, Hugging Face, and Goodfire partner on AI safety

Base Labs, a research group established by Baseten, has partnered with Hugging Face and Goodfire to develop safety evaluation and monitoring infrastructure for open-weight AI models. The initiative aims to create standardized methods for training and monitoring models to ensure safety without compromising their open nature.

The collaboration addresses growing concerns regarding the security of open models, particularly the rise of ‘abliteration’—a technique used to remove safeguards from models. Hugging Face currently hosts over 6,000 such abliterated models. The partnership leverages the specific strengths of each participant: Base Labs will focus on training and monitoring methodologies, Hugging Face provides model distribution via its hosting hub, and Goodfire contributes expertise in AI interpretability to explain model decision-making processes.

Baseten intends to integrate these research outputs into its existing deployment infrastructure to provide live runtime monitoring as a managed service. The companies are framing this work as a transparent standard built directly into the training and deployment lifecycle rather than as an afterthought.