Base Labs partners with Hugging Face and Goodfire on AI safety initiative

Base Labs, the research group that AI infrastructure company Baseten spun up earlier this year, is joining forces with Hugging Face and Goodfire to develop and publish methods for training and monitoring open models. The collaboration puts three organizations with very different specialties, inference infrastructure, model hosting, and interpretability, on the same page about a problem that’s gotten increasingly hard to ignore: making open AI models safer without making them less open.

The initiative arrives at a moment when the AI industry is grappling with a series of safety incidents involving autonomous agents in mid-2026, events that have pushed the conversation about open-model security from academic curiosity to boardroom priority.

What the partnership actually involves

Base Labs will focus on developing and publishing methodologies for training open models with stronger safety properties and for monitoring those models once they’re deployed in production.

Baseten plans to integrate the research outputs directly into its existing deployment infrastructure, enabling live runtime monitoring of AI models running on Baseten’s platform as a managed service.

Hugging Face brings its role as the de facto hub for open-weight model distribution. Baseten became an official Hugging Face Inference Provider in August 2026, which means the two companies already have plumbing connecting their systems.

Goodfire, meanwhile, contributes its expertise in AI interpretability. The company raised $150 million in a Series B funding round in February 2026, reaching a $1.25 billion valuation.

Why open-model safety matters now

A notable incident involving Hugging Face in July 2026 highlighted vulnerabilities in the open-model ecosystem specifically. Nvidia led the formation of the Open Secure AI Alliance, signaling that even hardware companies see safety as foundational infrastructure. Base Labs’ partnership with Hugging Face and Goodfire fits into this same pattern.

The tension at the heart of this work is real and worth understanding. Open-weight models derive their value from transparency: anyone can inspect, modify, and deploy them. But that same openness creates attack surfaces that closed models don’t have. Base Labs is framing its goal as building actionable controls that preserve openness.

The competitive landscape for AI safety infrastructure

Regulated sectors don’t just prefer safety features. They require them. A hospital deploying an AI diagnostic tool needs to demonstrate ongoing monitoring and interpretability to satisfy regulators. A financial institution using AI for risk assessment faces similar scrutiny. By baking safety into its managed infrastructure, Baseten is essentially building compliance tooling directly into its product.

The $150 million that investors poured into Goodfire at a $1.25 billion valuation tells a similar story from the capital allocation side. Interpretability now commands unicorn-level funding because the market recognizes that understanding what models are doing internally is prerequisite to trusting them in high-stakes environments.