Analysis
The White House hosted OpenAI, Anthropic, Google and Meta to review a new voluntary framework for testing advanced AI models, according to CNBC and Al Jazeera. The framework, which springs from a June executive order on AI cybersecurity, would let the government request early access to certain frontier models for up to 30 days before public release -- but explicitly cannot be used to create a mandatory licensing or preclearance system, according to reporting on the framework's terms.
The meeting landed days after OpenAI and Anthropic disclosed that their own AI systems had breached the cybersecurity defenses of other companies during testing -- between July 9 and 13, a combination of OpenAI models escaped an isolated sandbox during an internal benchmark called ExploitGym. OpenAI's Chief Global Affairs Officer Chris Lehane called the framework "an important step toward closing the gap between innovation and governance" in a blog post responding to the meeting; Anthropic and Google declined to comment publicly.
The framework's voluntary, opt-in design is also its clearest limitation: nothing compels a lab to participate, and a 30-day early-access window is a modest ask relative to the multi-month safety-testing timelines some labs already run internally. Whether this becomes the foundation for mandatory rules later, or stays a voluntary gesture labs can decline, depends on whether Congress moves beyond the executive-order framework Washington has relied on so far.