Potential misuse of AI by ‘bad actors’ a worry, says Google DeepMind CEO Demis Hassabis

Demis Hassabis, chief executive officer (CEO) of Google DeepMind, on Wednesday warned that risks related to biosecurity and cybersecurity need urgent attention as artificial intelligence systems become more powerful.

“I think we need to worry about things like bio and cyber risk in AI very soon. The current systems are getting pretty good at cyber,” Hassabis said. “You need to make sure that the defences are stronger than the offences.”

Speaking at the AI Impact Summit in New Delhi, Hassabis said artificial general intelligence (AGI) is now close and could arrive in the next five to eight years. “We’re starting to see what these systems can do,” he said.


Hassabis said that as AI systems become more powerful, there is growing concern around misuse by "bad actors." These could include “individuals, groups, or even nation-states who might repurpose these dual-use systems for harmful ends”.

He also flagged risks emerging as the world moves closer to AGI and into what he described as an agentic era, where AI systems become more autonomous. “I think we’ll see a lot of this happening this year and next year,” he said.

In such a scenario, Hassabis said it would be critical to put guardrails in place to ensure systems behave as intended and do not drift into unintended or unsafe areas. He said one challenge is societal, which will require international dialogue and ideally a minimum set of global standards. The other is technical — ensuring AI systems are robust, reliable and aligned.

Hassabis, the creator of AlphaFold, an AI system that solved a 50-year-old problem, said one of the biggest gaps in today’s AI systems is continual learning. “Today, systems are trained, fine-tuned, and then essentially frozen before being released into the world,” he said.

According to Hassabis, AI systems should be able to learn from real-world experiences, adapt to different situations and personalise themselves to the tasks they are given — "something current systems are unable to do."

Hassabis said another limitation is long-term planning. While AI models can plan over short periods, they struggle to make long, coherent plans over months or years, the way humans do.

He also pointed to a lack of consistency in current systems. “Today’s systems are what I would call ‘jagged intelligences’,” he said, explaining that they can perform some tasks extremely well but others badly.

Looking ahead, Hassabis said AI will significantly increase the productivity of scientists and experts. He added that cross-disciplinary research, which requires understanding multiple fields at the same time, will benefit the most, as AI helps researchers identify connections that are otherwise difficult to spot.