Frontier safety at Google DeepMind

Google DeepMind describes a safety approach for advanced foundation models centered on its Frontier Safety Framework and Google’s AI Principles. The framework identifies model capability tiers, checks for those capabilities across a model’s lifecycle, and prepares measures to address severe risks when they emerge. External contributors may be involved when needed to inform the process. The company says the framework will continue to change as new risks are identified, and lists versions from 2024 through April 2026.