Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire
News Source
•Thu, 17 Sep 2026 17:15:59 +0000
📰 What Happened
Baseten has launched a new safety standard through its research group, Base Labs. It is working with Hugging Face and Goodfire AI to build tools that test and keep watch over open-weight AI models. Open-weight models are ones anyone can download and change.
The problem is growing. Hugging Face, which hosts open source AI models, lists more than 6,000 “abliterated” models. Those are open models whose safety limits have been removed using a method called abliteration.
🔍 The Backstory
Open models are free to download, so anyone can strip out the rules that stop them doing harmful things. Closed models like ChatGPT run on the company’s own computers and cannot be changed by users.
Base Labs says it will publish ways to train and check open models. The partners argue that safety should be built into a model from the start, not added later. Goodfire specialises in opening up the “black box” of AI to show how models decide. Baseten raised $1.5 billion in June.
🎯 Why It Matters
Free, unfiltered AI models can be used to build scams and fake content. Better testing helps keep those tools from spreading.
Baseten has launched a new safety standard through its research group, Base Labs. It is working with Hugging Face and Goodfire AI to build tools that test and keep watch over open-weight AI models. Open-weight models are ones anyone can download and change.
The problem is growing. Hugging Face, which hosts open source AI models, lists more than 6,000 “abliterated” models. Those are open models whose safety limits have been removed using a method called abliteration.
Open models are free to download, so anyone can strip out the rules that stop them doing harmful things. Closed models like ChatGPT run on the company’s own computers and cannot be changed by users.
Base Labs says it will publish ways to train and check open models. The partners argue that safety should be built into a model from the start, not added later. Goodfire specialises in opening up the “black box” of AI to show how models decide. Baseten raised $1.5 billion in June.
Free, unfiltered AI models can be used to build scams and fake content. Better testing helps keep those tools from spreading.