Hassabis suggested new AI models should first be tested in offline sandboxes after other companies' models escaped closed evaluation environments.
New AI models should likely first be tested in offline "sandboxes," after other companies' models escaped the closed environments used to assess their capabilities, he added.