FT: Frontier AI Labs Race to Tighten Biological-Risk Safeguards as Bio Evals Prove Harder Than Cyber

The Financial Times reports OpenAI, Anthropic and Google DeepMind are stepping up biological-risk testing of frontier models, coordinating through the Frontier Model Forum and Sentinel Bio on a multi-participant BSL-2 wet-lab study to measure whether AI assistance boosts novice success on real biosecurity tasks. Labs concede bio evals are harder than cybersecurity tests, where capabilities can be probed in fully digital environments.