As we build more capable and more personalized AI, reliability, security, and user protections are more important than ever.

Advanced models require an advanced approach to safety — one that scales with the technology. Today, we’re detailing that work: our updated Advanced AI Scaling Framework, our Safety & Preparedness Report for Muse Spark, and new advances in how our models reason about safety from the ground up, so that as our AI becomes more capable, our protections keep pace.

Today, we’re building on our original Frontier AI Framework and publishing a significantly updated and more rigorous version: the Advanced AI Scaling Framework. This update broadens the types of risks we evaluate, strengthens how we make deployment decisions, and introduces new Safety & Preparedness Reports. More specifically, this Framework outlines how we identify and assess the most severe and emerging risks, including chemical and biological, cybersecurity, and a new section to evaluate risks around loss of control. As models become more advanced, we’re evaluating how they perform when given greater autonomy and whether the controls around that behavior work as intended. These standards apply across our frontier deployments, whether they’re open, controlled API access, or closed models.

In practice, this also means mapping potential risks, evaluating models before and after safeguards are applied to confirm they work in the real world, and only deploying models when they meet the standards set by our Framework. For people who use Meta AI across our apps, this means the models powering their experience have been evaluated across a broad spectrum of risks before we make them available.

While our updated Framework strengthens the standards and safeguards for our most capable models, our new Safety & Preparedness Reports will show how we are meeting them. These reports will detail our risk assessments, evaluation results, the rationale behind our deployment decisions, and any limitations we’re still working to address. This transparency means we will share what we found, how we tested our models, where our evaluations fell short, and how we closed those gaps.