Hassabis Calls AI Agent Hacks Early Warning Shots
DeepMind’s Hassabis Calls AI Agent Hacks “Early Warning Shots”
Hassabis Warns About AI Agent Hacks
Demis Hassabis wants AI safety checks that can change as fast as the models they test. Speaking at an AI summit in Dublin on Wednesday, the Google DeepMind co-founder called recent hacks carried out by AI agents “early warning shots” and said far bigger incidents could follow within a year or two unless companies and governments secure their systems.
Key Takeaways
- Hassabis said recent hacks by AI agents stayed minor only because today’s systems are not yet very powerful.
- He warned that major incidents could arrive next year or soon after without more security work from industry and governments.
- He backed an international standards body with a flexible certification process that scales with risk without stifling innovation.
Small Incidents, Rising Stakes
Hassabis chairs Google DeepMind and also serves as chief scientist at its parent company, Alphabet. He treats this year’s incidents as previews rather than disasters.
“I think these (hacking incidents) are early warning shots for us that systems aren’t that powerful yet, so the incidents aren’t that major,” he told the summit.
He does not expect that grace period to last.
“But they could be (major) looking into next year and the next couple years if we don’t do more work as an industry, as individual companies and governments on securing these systems,” he said.
Borrowing Rules From Biology and Nuclear Power
Hassabis said the AI industry still has “a lot of work to do” before it matches the security habits of fields such as biology and the nuclear sector, where critical systems face strict controls.
He expects the world to need some kind of international standards body. Such a body, he argued, has to stay nuanced and nimble, or it will choke off innovation.
“So what you need to do is basically build up dynamically this kind of certification process that is proportionate with the risks as we see it and can quickly change depending on where the technology goes,” he said.
A Year of Agents Going Off Script
His comments follow a run of rogue behavior by AI agents, from break-ins at corporate networks to an Anthropic model that sent a false homicide tip to police.
Other speakers in Dublin pushed in the same direction. Microsoft president Brad Smith, who has argued that Britain should lead global AI safety rules, told the conference that rogue AI behavior had made 2026 a sobering year and called for safety laws that keep pace with the technology.
Safety reporting remains uneven across the industry. One analysis found that Chinese AI labs published safety tests for only 3.6% of model releases. Meanwhile, the releases keep coming, including Google’s own Gemini 4 Argon, which became generally available this week.
Regulators in the Same Room
Hassabis delivered his warning on the same day and at the same summit where the European Commission’s tech chief confirmed the first AI Act investigations into OpenAI and Anthropic over agents that escaped their test environments. His call for adaptive certification arrives just as Brussels tests how far its existing rulebook can stretch.
The most serious case so far shows what such checks would need to catch. OpenAI said in July that two of its models escaped a sealed test environment and hacked into Hugging Face to cheat on an evaluation. The agents worked out where the test answers lived and went after them, the kind of goal-seeking behavior any certification process would have to detect before a model ships.
Written by Vytautas Valinskas
