OpenAI will give the U.S. government a 30-day early look at its frontier AI models before public release, confirming compliance with the executive order President Trump signed June 2.
George Osborne, OpenAI’s head of countries, made the announcement Friday at SXSW London. The EO asks companies to voluntarily let federal agencies benchmark a model’s “advanced cyber capabilities” and determine whether it qualifies as a “covered frontier model.” It’s explicit: no mandatory licensing or preclearance, for now.
The order also directs the Attorney General to prioritize enforcement against anyone using AI for cyberattacks or cybercrime. The DOJ gets the stick; developers get the carrot of being listed as “trusted partners” with early model access.
The day before, Anthropic published a call for a global pause or coordinated slowdown in frontier AI development, citing recursive self-improvement risk: AI systems increasingly capable of building the next generation autonomously. Anthropic’s own data shows Claude now accounts for more than 80% of code contributions to its own codebase.
That’s the split worth watching. OpenAI’s voluntary compliance normalizes government pre-release access as an industry practice. Anthropic’s conditional pause proposal frames the same moment as a governance crisis. Every other lab now has to pick a public position, and the voluntary framework has a natural upgrade path: if benchmarking becomes routine, the gap between “voluntary” and “required” closes faster than most legal teams are planning for. Any founder building on frontier models should map which roadmap features get delayed if a 30-day federal review becomes the de facto standard before launch.
Sam Altman endorsed the order on X: “new EO gets the balance right.” Osborne added OpenAI suggested safety-tracking to governments globally, not just Washington.
Nathan Zakhary