Demis Hassabis Says AI Hacks Could Signal Larger Security Risks

Google DeepMind Chairman Demis Hassabis said recent hacking incidents involving autonomous AI agents were an early sign of security risks that could grow as AI advances. Speaking at the International AI Summit 2026 in Dublin on Oct. 14, he urged companies and governments to strengthen defenses and test models in offline sandboxes before release.

“The incidents aren’t that major,” Hassabis said, describing them as “early warning shots.” Without additional security work, he said, problems could become more serious “looking into next year and the next couple of years.”

AI agents can take actions on a user’s behalf. That can make them more useful, while also increasing the consequences of mistakes or misuse. Hassabis called for using AI systems to identify and fix software vulnerabilities before products are released.

He also urged developers to draw on safety expertise from fields such as biology and nuclear technology to help contain risks. Hassabis framed stronger security as a way to protect AI’s benefits while development continues.

Microsoft President Brad Smith proposed AI safety standards, mandatory incident reporting and controls that users could activate to stop agents. “There are more safety rules for a carton of milk than for AI,” Smith said.

Hassabis did not identify the hacking incidents he referenced. The remarks described at the summit did not specify their number, targets or effects.