Base Labs Launches Open-Weight AI Safety Partnership With Hugging Face

Base Labs, the research division of AI inference platform Baseten, has partnered with Hugging Face and Goodfire AI to establish a new safety infrastructure standard for open-weight artificial intelligence models. Announced on Wednesday, the collaboration addresses growing industry concerns regarding the vulnerability of open-source models to a technique known as abliteration, which allows users to strip away safety guardrails. The threat is already pervasive, with Hugging Face currently cataloging over six thousand abliterated models on its repository. Rather than treating safety as an afterthought, the partnership aims to embed protective measures directly into the training and deployment pipelines of open models. Base Labs intends to develop and publish transparent methodologies that will serve as an industry standard, emphasizing that openness inherently improves AI safety by increasing visibility into model behavior and accelerating the translation of research into actionable controls. Goodfire AI, which specializes in model interpretability, will contribute technical expertise to ensure safety mechanisms are intrinsically woven into model architecture. Hugging Face will provide the hosting and distribution infrastructure necessary to scale these standards across the developer community. The initiative coincides with significant capital inflows into the AI infrastructure sector. Baseten recently secured a 1.5 billion dollar Series F funding round, elevating its valuation to 13 billion dollars, while Goodfire AI previously raised 150 million dollars in Series B financing to expand its interpretability platform. Both organizations are leveraging their resources to professionalize open-model safety protocols. Moving forward, Base Labs has issued an open invitation to the broader developer ecosystem to contribute to the framework. The companies aim to cultivate a collaborative environment where open-weight models remain both accessible and rigorously monitored. By shifting the safety paradigm from reactive patching to proactive architectural integration, the partnership seeks to establish a reproducible benchmark for the open-source AI community, reducing the attack surface of publicly available models while preserving the collaborative advantages of open development.