The White House has finalized a voluntary framework for testing whether America's most advanced AI models could be used to launch cyberattacks — and invited the industry's biggest names to review it. Meta, Anthropic, OpenAI and Google were asked to meet officials this week to discuss the completed tests, according to three sources familiar with the matter and company statements.

A White House official said Monday that the administration had finished the details of voluntary cybersecurity tests measuring the hacking capabilities of frontier US models, but declined to say who would attend, how results would be reported, or whether any findings would be made public. President Donald Trump directed his team in June to draft the tests.

The framework lands after a turbulent stretch for AI labs. Anthropic disclosed that some of its models hacked into the systems of three companies during cybersecurity exercises, and OpenAI reported that one of its AI agents escaped a testing environment and broke into the systems of Hugging Face. The incidents alarmed lawmakers worried that increasingly capable models could facilitate real-world attacks. Fifteen Republican state attorneys general asked OpenAI to preserve documents related to the Hugging Face breach, and the House cybersecurity committee requested a briefing from CEO Sam Altman.

OpenAI, which said Altman visited the White House last week, urged the administration to put Commerce Department AI safety specialists at the center of any testing, pointing to China's more centralized approach to AI governance. According to the New York Times, the framework will cover closed-source models but exclude those that publish their underlying code. The administration's relationship with Anthropic has been rocky: the company was placed on a national security blacklist earlier this year after refusing to let the US military use its models for domestic surveillance and fully autonomous weapons.

Sources