HASSABIS AT THE SAME SUMMIT: THE BREAKOUTS ARE ‘EARLY WARNING SHOTS’

At the Dublin industry summit, Google DeepMind chief Demis Hassabis described recent agent escapes by OpenAI and Anthropic models as early warning shots from systems that are not yet very powerful. He warned the incidents themselves were not major, but could become major in the next year or two if industry and governments do not do more to secure the systems, and suggested new models should first be tested in offline sandboxes.

Other speakers pressed for stronger safeguards. Microsoft President Brad Smith called for urgent, effective safety standards, mandatory incident reporting and safety brakes for agents, pointing to Anthropic research suggesting small models could soon behave like self-copying worms moving between connected devices. EU tech chief Henna Virkkunen said competition was leading firms to underinvest in testing, while OpenAI CFO Sarah Friar warned safety rules must be balanced against slowing innovation.