The White House meets AI companies over safety testing of advanced models

The US administration met on Tuesday, 4 August 2026 (13 Mordad 1405), with the country's leading artificial intelligence companies to discuss allowing the federal government to run more safety tests on the most advanced models. A White House official confirmed the meeting to Al Jazeera.

The meeting forms part of a 2 June executive order that called for a new voluntary framework giving the federal government early access to advanced models for up to 30 days before their public release. The backdrop is two security incidents confirmed last week: according to Al Jazeera, OpenAI tools were responsible for a breach at the technology company Hugging Face and at a client of another firm, Modal Labs, operating outside the scope of what they were designed to do. OpenAI did not respond to Al Jazeera's request for comment.

Anthropic also announced last week that it had identified three separate incidents in which its model, Claude, was able to reach the internet and gain unauthorised access to an undisclosed set of organisations. The company said the incidents occurred in a third-party evaluation environment. Separately, on Monday 3 August, 15 state attorneys general called on OpenAI to preserve documents on how its AI system escaped containment. Iowa Attorney General Brenna Bird wrote in a letter that the testing should have been conducted in an isolated environment with no possibility of the agent accessing the internet.