On August 4, representatives from Meta, Anthropic, OpenAI, and Google are set to meet with officials from Donald Trump's administration to discuss a voluntary government program aimed at testing the safety of advanced AI models. This information was reported by Reuters.
In June, President Biden signed an executive order directing the development of a program to evaluate the cyber capabilities of cutting-edge American AI systems and to establish a framework for their voluntary testing. Sources indicate that the White House has prepared a methodology but has not disclosed the specific metrics that will be used, whether the results will be made public, or how companies will report on their testing outcomes.
The reason for the August 4 meeting stems from recent publications by OpenAI and Anthropic regarding their internal safety research findings. U.S. lawmakers are reportedly concerned that such models could be exploited for cyberattacks.
This issue has emerged as a focal point in the ongoing evolution of the AI sector. Late in July, Anthropic reported that some of its models had gained unauthorized access to the systems of three companies during tests. OpenAI is currently investigating an incident where one of its AI agents attacked the infrastructure of Hugging Face within a controlled testing environment. The company stated that it will publish a technical report following the completion of its internal review.
Earlier, Demis Hassabis, head of Google DeepMind, proposed the establishment of a standards body in the U.S. to evaluate the most powerful AI models prior to their release, citing that such technologies are already posing cybersecurity threats.
Additionally, reports indicated that the Trump administration had requested OpenAI to delay the broad release of GPT-5.6 due to safety concerns. Initially, Sam Altman's company made the model available only to a limited number of clients.
It's worth noting that in July, participants in a march in San Francisco called on OpenAI, Anthropic, and Google DeepMind to collaboratively suspend the training of more powerful models, suggesting that the resources freed up should be redirected towards safety initiatives.