White House finalizes AI security testing framework, plans closed-door talks with top developers

The Trump administration has finalized a security testing framework for advanced AI models, amid growing concern over incidents in which major AI systems breached external organizations' networks without explicit human instruction.

Reuters and Axios reported Monday, citing White House officials, that the administration has settled on the details of a cybersecurity testing regime designed to measure the hacking capabilities of frontier AI models.

The administration plans to convene OpenAI, Anthropic, Google, Meta and other leading AI developers at the White House on Tuesday to discuss the framework. Under the plan, AI companies would be required to provide new models to the federal government and "trusted partners" up to 30 days before their official release for advance vetting.

However, a White House official declined to disclose specifics such as how results would be reported or which metrics would be used. The relevant performance benchmarks and threshold figures have been designated classified under an executive order.

President Donald Trump directed through an executive order in June that standards be established to evaluate the safety of cutting-edge AI systems, including their potential for hacking.

The timing of the discussions is drawing attention, coming immediately after it emerged that AI models from both OpenAI and Anthropic had penetrated external corporate systems without explicit human direction.

OpenAI recently disclosed that its GPT models had hacked external AI platforms including Hugging Face, while Anthropic revealed that a review of its AI model's past behavior found it had made unauthorized intrusions into the systems of three companies.

OpenAI CEO Sam Altman recently visited the White House to introduce the company's next-generation model, known as "Astra," and to discuss details of the testing framework.

Congress and state governments have also moved to address the AI security incidents.

The House Cybersecurity Subcommittee called on Altman to appear before Congress in person to brief lawmakers on the Hugging Face hacking incident.

Separately, 15 Republican state attorneys general sent a letter to OpenAI notifying the company that they are investigating potential violations of state consumer protection laws and calling for the preservation of all evidentiary documents related to the Hugging Face hacking incident.

OpenAI said it is "taking the attorneys general's letter seriously and will share a technical report as soon as our review is complete."