The United States government established an early access program with OpenAI and Anthropic to review model vulnerabilities prior to public release