Baseten Aims for a Security Standard for Open Models

Key Takeaways

1Over 6,000 abliterated models are listed on Hugging Face

2Baseten, Hugging Face, and Goodfire are launching a partnership for an integrated security standard

3The technical details of the partnership are yet to be clarified

4Baseten is calling on developers to contribute to the framework

Why it matters : The initiative aims to enhance the security of open-weight models, as their vulnerability to abliterations is a major concern.

In response to the proliferation of open-weight models that have been abliterated, Baseten is partnering with Hugging Face and Goodfire to promote an integrated security framework. The partners advocate for openness as an asset, while leaving technical details pending and inviting the developer community to contribute.

Massive Abliteration and Unclear Modalities

Hugging Face lists over 6,000 abliterated models, highlighting the scale of the problem posed by the removal of protections on open-weight models. This technique, known as abliteration, is identified as a means to render these systems dangerous. The partners have yet to detail the technical workings of their collaboration. Meanwhile, Baseten is inviting developers to participate in the development of the framework. The company claims it aims to build a safe and accessible ecosystem of open models for everyone.

Players with Financial and Technical Resources

In June, Baseten, a company specializing in AI inference, raised $1.5 billion in a Series F funding round, bringing its valuation to $13 billion. Goodfire, which focuses on the interpretability and explanation of model decisions, secured $150 million in a Series B round led by B Capital earlier this year. Hugging Face hosts open-source AI models. These financial and technical resources position the partnership within an already structured ecosystem.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

An Integrated Framework for Training, Deploying, and Monitoring

Baseten has launched, through its research branch Base Labs, a security infrastructure standard in partnership with Hugging Face and Goodfire AI. The goal is to build an evaluation and monitoring infrastructure for open-weight models. Baseten presents this project as a transparent standard integrated into the training and deployment of models, rather than an afterthought. Base Labs, established earlier this year, plans to develop and publish methods for training and monitoring these models.

Openness Claimed as a Security Asset

Baseten asserts that openness represents an advantage for AI security. According to the company, it allows for better visibility into model behavior and facilitates the transformation of security research into actionable and transparent controls, unlike closed environments. Goodfire shares this approach, believing that security must be integrated into open models and ensured by those who serve them.