“AI models are expected to reach AI Safety Level 3 (ASL-3), where they can enhance the capabilities of non-state actors in dangerous domains, in 2025.”