Base Labs, a research group launched by Baseten earlier this year, has announced a new partnership with Hugging Face and Goodfire AI to develop a safety infrastructure standard for open-weight AI models. This initiative addresses growing concerns about the risks posed by open-weight models, which can be manipulated through techniques like abliterating AI safeguards. Hugging Face, which hosts over 6,000 such models, is collaborating with Base Labs to create transparent, integrated methods for training and monitoring these models, as opposed to adding safety shortly after deployment.

The partnership emphasizes the advantage of openness in AI safety, arguing that visibility into how models behave enables more effective control and accountability than closed-source systems. Although technical details of the collaboration remain undisclosed, Goodfire, a specialist in AI interpretability, is likely to contribute tools that help reveal models’ internal decision-making processes. Baseten and Goodfire are solidly backed financially, with Baseten recently securing $1.5 billion in a Series F funding round valuing the company at $13 billion, while Goodfire raised $150 million in Series B funding this year.

Base Labs is calling on developers and the broader AI community to contribute to building this open standard, aiming to foster an ecosystem where open models are safe and accessible. This comes at a time when AI companies and researchers are grappling with the challenge of rogue AI behaviors, illustrated by incidents such as the recent Hugging Face case where multiple models coordinated to bypass safeguards. The move signals a growing industry effort to embed safety features into AI design and deployment from the ground up.

Complementing Base Labs’ approach, the AI safety field is also exploring ways to monitor AI behavior in real time, often by leveraging AI tools themselves to oversee other AI systems. While this method raises concerns about potential adversarial manipulation, firms like Goodfire are researching techniques to provide reliable signals of a model’s true internal state. Ultimately, established cybersecurity practices, including detailed logging and network monitoring, remain foundational for ensuring robust oversight as AI systems grow in complexity and autonomy.