LLMs

The Government Just Told OpenAI When It Can Ship. That Changes Everything.

The White House asked OpenAI to delay the public rollout of GPT-5.6, and OpenAI complied, restricting the model to a select group of partners rather than releasing it broadly. OpenAI pushed back rhetorically, stating that government access controls 'keep the best tools from users, developers, enterprises, cyber defenders, and global partners who need them,' but the compliance itself is the signal that matters. Observers like Ethan Mollick are asking the harder question: what specific risks are driving this? If the concern is military implications, that's one category of problem. If it's information security vulnerabilities that could be exploited at scale, as some Five Eyes signals suggest, the implications are far broader, especially as open-weight models of comparable capability are expected to arrive within 6 to 12 months regardless of what any government requests. The policy vacuum is the real danger. Without clarity on what 'Mythos-class' risk actually means, companies and governments alike can't prepare defenses before open-source equivalents make the question moot. Meanwhile, Anthropic is caught in a parallel bind: its own vocal warnings about advanced AI dangers may have made it a target for export restrictions, a cautionary tale about how safety rhetoric can become a regulatory liability. The era when AI labs set their own release calendars, on their own terms, appears to be over.