“Three leading AI developers applied enhanced safeguards to their new models because internal pre-deployment testing could not rule out the possibility of misuse for creating biological weapons.”