Scaling How We Build and Test Our Most Advanced AI

Meta has announced an updated safety approach for its most capable models. The company says it is publishing a significantly revised Advanced AI Scaling Framework that builds on its earlier Frontier AI Framework, a new Safety and Preparedness Report covering its Muse Spark model, and new work on how its models reason about safety.

The framework widens the risks Meta evaluates. It covers chemical and biological threats and cybersecurity, and adds a new section on loss of control, which examines how models behave when given more autonomy and whether controls around that behavior work. The standards apply to frontier deployments whether open, accessed through controlled API or closed. In practice Meta says it maps risks, tests models before and after safeguards, and deploys only those meeting the framework's standards. The new reports are meant to show how the standards are met, setting out risk assessments, results, the reasoning behind deployment decisions and limitations still being addressed.

For Muse Spark, Meta says it tested before and after protections, against serious risks and against long-standing policies on violence, child safety, criminal wrongdoing and ideological balance. It describes thousands of scenarios designed to find weaknesses and live monitoring to catch unexpected problems. Meta reports strong safeguards across measured categories, says its evaluations placed Muse Spark at the frontier when it comes to steering clear of ideological bias, and says it lacks the autonomous capability needed to pose control risks.

Meta also describes a change in method. Rather than training models on specific scenarios one by one, it turned its trust and safety guidelines into testable principles and trained the model on the reasons behind them, so it can handle unanticipated situations. Meta says human teams still design and validate the principles and add further guardrails.