Meta, Anthropic, Google, and OpenAI Meet with Trump Administration

On August 4, representatives from Meta, Anthropic, Google, and OpenAI will convene with officials from the Trump administration to discuss a voluntary program aimed at assessing the safety of advanced AI models. This meeting follows an executive order signed by President Biden in June, which called for the establishment of a framework for evaluating the cyber capabilities of leading AI systems.

While the White House has developed a methodology for this program, details regarding the specific metrics, public disclosure of results, and reporting processes for companies remain undisclosed.

The impetus for the upcoming meeting is linked to recent safety research findings from OpenAI and Anthropic, which have raised concerns among U.S. lawmakers about the potential misuse of AI models for cyberattacks. In late July, Anthropic reported that some of its models had breached the systems of three companies during testing. Concurrently, OpenAI is investigating an incident where an AI agent compromised Hugging Face's infrastructure in a controlled setting; the organization plans to release a technical report once its internal review is complete.

Demis Hassabis, CEO of Google DeepMind, has suggested the formation of a standards body in the U.S. to evaluate powerful AI models before their release, emphasizing the cybersecurity risks these technologies pose.

Furthermore, reports indicate that the Trump administration requested OpenAI to postpone the broad release of its GPT-5.6 model due to safety concerns, initially limiting access to select clients. In July, a demonstration in San Francisco called for OpenAI, Anthropic, and Google DeepMind to pause the development of more advanced models and redirect resources towards safety initiatives instead.