White House Finalizes Advanced AI Security Verification Plan, to Discuss Privately with AI Companies

Amid security incidents where major artificial intelligence (AI) models breached external institutional systems, the U.S. administration under Donald Trump has reportedly finalized a related security testing plan.

Local time on the 3rd, Reuters and U.S. online media outlet Axios reported, citing a White House official, that the U.S. administration has finalized details for cybersecurity tests to measure the hacking capabilities of cutting-edge AI models.

The Trump administration plans to invite major AI developers, including OpenAI, Anthropic, Google, and Meta, to the White House on the 4th to discuss the matter.

This testing plan includes a structure where AI developers provide their models early to the federal government and trusted partners for up to 30 days prior to the official release of new models to undergo verification.

However, the White House official did not disclose detailed information, such as the reporting method for the results or what metrics will be used in the test.

This is because the relevant performance benchmarks and threshold figures have been designated as classified under an executive order.

President Trump directed the establishment of safety evaluation test standards, including the potential for hacking by cutting-edge AI systems, through an executive order last June.

This discussion on AI security is drawing attention as it comes right after revelations that AI models from OpenAI and Anthropic breached external corporate systems without explicit human instructions.

OpenAI recently stated that its GPT models hacked external AI platforms such as Hugging Face, and Anthropic also recently disclosed that a review of its AI model's past behavior revealed unauthorized breaches into three corporate systems.

(Photo: Yonhap News)
※ Please note: This article was translated by AI and may contain errors.