This week in AI, the spotlight is on AI-powered biology, scientific breakthroughs, and stronger AI safety. Anthropic is improving biology safeguards around its Fable 5 model, reducing unnecessary refusals while continuing to restrict potentially harmful biological assistance. The company says its updated safety classifier has reduced unnecessary biology-related refusals by around 85%, allowing the model to respond more effectively to legitimate scientific and educational questions. Researchers at Stanford University and the Arc Institute used AI to design entirely new viral genomes. They synthesized 16 AI-designed genomes, and the resulting viruses were able to infect bacteria. The AI-generated viruses were based on Phi X-174, a naturally occurring bacteriophage that infects bacteria rather than humans. OpenAI is strengthening safeguards as its AI systems approach more advanced cybersecurity capabilities, expanding monitoring, tightening access to networks and tools, and introducing additional controls for high-risk activities. Together, these developments highlight a critical shift in AI: as models become capable of contributing to advanced scientific and technical work, building effective safety systems alongside those capabilities is becoming increasingly essential.