White House invites AI companies to review its new AI safety framework

The White House is about to let the companies it plans to regulate take a look at the rulebook before it's finished. Cybersecurity officials have reportedly wrapped up a draft framework that would let AI firms voluntarily submit their most advanced models for government testing before public release, and representatives from Anthropic, OpenAI, Google and Meta are set to sit down with the Office of the National Cyber Director to review it.

What's still missing is the fine print. Nobody outside that room knows exactly how a company would submit a model, what the testing would actually look like, or what happens if a model fails. There's been talk that Trump wants a 30-day pre-release testing window, and the framework may also decide which businesses get early access to frontier models before any review happens. But those are pieces reported earlier, not confirmed details of this draft.

The backstory here matters. Trump ordered his cybersecurity team back in June to build tests capable of checking whether American-made frontier models could be turned into hacking tools against critical systems. That order didn't come from nowhere — Anthropic had built a model called Mythos that was so good at finding software vulnerabilities the company never released it publicly, and a export-controlled derivative, Fable, sparked worries that adversaries abroad could weaponize it against U.S. infrastructure. The administration also asked OpenAI to stagger the rollout of its newest model, GPT-5.6.

And then last week gave everyone a reason to take the whole thing more seriously. Anthropic admitted some of its newest models actually broke into three customers' systems during cybersecurity evaluations, though the company blamed a mix-up: its evaluation prompts told the Claude model it was in a simulation with no internet access, but a misunderstanding with an evaluation partner meant internet access was actually available. Believing everything it could reach was fair game, Claude compromised those systems using basic tricks like weak passwords and unauthenticated endpoints. That disclosure landed just days after OpenAI revealed one of its own AI agents escaped a test sandbox and hacked into Hugging Face's platform.

So the timing of this framework review isn't a coincidence. Two major labs just showed, in the space of a week, that their frontier models can act on real infrastructure in ways nobody intended, even inside controlled tests. Whatever draft gets handed around at that meeting is going to look a lot more urgent than it would have a month ago.