Meta

Advanced AI Scaling Framework (April 8, 2026). Replaces and supersedes the original Frontier AI Framework. Broadens risk categories to add a new section on loss of control. Strengthens deployment decision criteria. Introduces Safety & Preparedness Reports as a public transparency artifact. Standards apply to all Meta frontier deployments, open or closed.

• ·Replaces and supersedes the original Frontier AI Framework
• ·Adds new section on loss of control risks
• ·Strengthens deployment decision criteria
• ·Introduces Safety & Preparedness Reports for each frontier model release
• ·Applies uniformly to open-weights releases AND controlled API access AND closed models
• ·First Safety & Preparedness Report published with Muse Spark

Original Meta Frontier AI Framework. Introduced two outcome tiers (Critical Risk, High Risk) tied to bio/chem and cybersecurity risk categories. Outcome-based rather than capability-threshold-based: Meta evaluates whether an AI's use could enable a defined catastrophic outcome rather than whether it crosses a capability bar.

• ·First public Meta safety framework
• ·Outcome-based: Critical Risk and High Risk outcomes defined per category
• ·Categories: biological weapons proliferation, cybersecurity attacks
• ·Decision on release based on whether marginal uplift to a threat actor crosses outcome thresholds