White House convenes top AI labs on safety testing

The White House brought OpenAI, Anthropic and Google together on Aug. 3 to talk through voluntary safety testing for frontier models, per CNN. Dario Amodei attended for Anthropic.

The meeting is the first concrete step out of June's executive order on AI cybersecurity, which set up an opt-in review process instead of binding rules. What gets tested, and who sees the results, stays up to each company.

It comes days after more than 1,200 senior staff across the major labs signed an open letter asking the government to build tools that can slow frontier development until safety work catches up. CNN doesn't say whether any of Aug. 3's attendees signed it, or what happens if a lab declines the voluntary review.

Voluntary compliance regimes have a weak track record. Companies tend to test what's convenient and skip what slows shipping. This time the pressure comes from inside the labs, not from outside critics or regulators. That's harder to dismiss than an op-ed, and it's worth watching whether this meeting produces anything the labs weren't already planning to do.