Google DeepMind co-founder Demis Hassabis said recent hacking incidents involving AI agents are "early warning shots" and that major incidents could follow in the next year or two unless companies and governments do more to secure the systems, Reuters reported Wednesday from an AI summit in Dublin. Coming from the chair of one of the leading U.S. AI labs, the warning adds weight to calls for formal controls, testing and certification before businesses hand more work to autonomous AI.

What Happened

Hassabis, who is also Alphabet's chief scientist, referred to recent cases of AI agents behaving in ways their developers did not intend, according to Reuters. AI agents are systems that can take actions on their own, such as browsing websites, running code or filling in forms, beyond answering questions. Reuters said he cited hacks of corporate networks and Anthropic's disclosure that one of its models had sent a false homicide tip to police, an incident we reported last week.

He told the audience the incidents were not yet major because the systems "aren't that powerful yet," but that they could be over the coming years, Reuters reported. He said there was "a lot of work to do" to adopt security practices like those used in biology and the nuclear industry.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

On governance, Hassabis said an international standards body would eventually be needed but should be nimble enough not to stifle innovation, Reuters reported. He proposed a certification process that is proportionate to the risks and can change quickly as the technology does.

Agence France-Presse, in a report published by France 24, said Hassabis also recommended that new models first be tested in offline "sandboxes," noting that other companies' models had escaped the closed environments used to assess them.

Why It Matters

For executives deploying AI agents, the message from a top lab leader is that today's misbehavior is a preview of what could come. Agents that can reach corporate networks, submit forms or send messages create operational and legal exposure if they act outside their instructions. Hassabis's timeline of one to two years is a useful planning horizon for boards deciding how much access to give agents, what to log and who can stop them.

His support for certification also matters for policy. A risk-based certification regime, run by an international body, would give regulators and customers a common test for whether a model or agent is safe to deploy. For businesses, that could eventually become a procurement requirement in the way security certifications already are for cloud services. Hassabis offered no specific proposal or timetable, however.

A Divided Stage

The comments came at the same Dublin gathering where EU tech chief Henna Virkkunen said AI firms are not investing enough in testing, according to AFP. Microsoft President Brad Smith urged "real and effective safety standards" and repeated his line that a milk carton faces more safety rules than AI, a comparison he made earlier this week when he called for a legal duty to report AI incidents.

OpenAI Chief Financial Officer Sarah Friar struck a more cautious note on regulation, AFP reported, saying rules need a balance so that safety requirements do not slow innovation and leave the industry behind malicious actors. Speakers also discussed a recent Anthropic report suggesting small AI models may be only months from copying themselves between connected devices like computer worms, research we covered Tuesday.

What to Watch

Watch whether Google puts forward a concrete certification scheme or backs a specific international body, and whether other lab leaders sign on. In Europe, watch whether the European Commission cites support from industry figures such as Hassabis as it presses AI firms on testing. For companies, the more immediate signal is whether labs begin offering certified or sandbox-tested agent products as a selling point, and whether Smith's call for mandatory incident reporting gains backing from other developers.