“Anthropic's early safety research focused on the potential for AI systems to become proficient in biology and biological weapons.”