By securing early access to AI models, U.S. ​administration aim to identify threats ranging from cyberattacks to military misuse before the tools are widely deployed.