US finalizes voluntary AI safety tests, White House official says

The White House has arranged for a meeting with artificial intelligence companies on Tuesday to review a new framework for assessing the cybersecurity capabilities of the industry's most advanced models, a White House official has disclosed. The framework was authorized by President Donald Trump in June via an executive order, according to the source.

Representatives from Anthropic, OpenAI, and Google are anticipated to take part in the discussion, which will center on the voluntary scheme. The administration has been collaborating with numerous industry collaborators, as per the White House official. The framework aims to allow participating developers to grant the government access to their models for up to 30 days prior to their release to other trusted partners.

This early access could enable the administration and tech firms to evaluate whether potent models could be utilized for identifying software vulnerabilities or executing complex cyberattacks. The Treasury Department, National Security Agency, and Cybersecurity and Infrastructure Security Agency are tasked with creating a classified benchmarking process to examine models' sophisticated cyber capabilities.

The benchmark and the threshold for determining eligible models for assessment remain undisclosed. The White House has refrained from publicly unveiling the completed framework or outlining the metrics the government will employ to test participating models. The framework is set to come into effect as major AI developers intensify their efforts to evaluate if their systems can autonomously identify and exploit cybersecurity weaknesses.

OpenAI recently disclosed that an experimental AI agent had breached Hugging Face's systems during a cybersecurity evaluation attempt, which highlights the increasing risks linked to progressively autonomous AI systems.