Google DeepMind's Demis Hassabis Warns AI Systems Are Not Healthy Yet in Dublin

Demis Hassabis, British AI scientist, Nobel Prize laureate, and Google DeepMind co-founder, addressed the International AI Summit in Dublin with a mixed message. While optimistic about AI's potential to accelerate science, he urged the industry, companies, and governments to prioritize security before releasing autonomous AI systems more widely. "The systems are not healthy yet," he warned, and that cannot be ignored.

Arm the Defenders First

Hassabis called for an asymmetric advantage on the defensive side of cybersecurity. He proposed using AI systems to fix software vulnerabilities before those systems are generally released, giving defenders a meaningful edge. "Asymmetrically power the defenders, trust the defenders," he said. The approach would mean deploying AI in defensive cybersecurity roles ahead of broader public availability.

He pointed to recent hacking incidents involving autonomous AI agents as evidence of existing risk, noting that they were not major but still illustrate why security work cannot wait. To address this, he called on industry and governments to build secure sandboxes for testing autonomous agents before release.

Science Optimism with Real-World Limits

Hassabis, who together with John Jumper won the Nobel Prize in chemistry for developing an AI model that predicts the structure of nearly all 200 million proteins, expressed genuine optimism about AI's scientific impact. He cited researchers gaining better understanding of antibiotic resistance and creating images of enzymes that decompose plastic. He also referenced OpenAI's September claim of solving the Navier-Stokes existence and smoothness problem, one of the Millennium Prize Problems. He noted that developing new mathematics along the way can be more valuable than the solution itself.

However, he acknowledged clear limits in physics, chemistry, and biology. AI still requires a real-world loop: forming a hypothesis, running an experiment, gathering data, and validating the result. "There will always be a kind of real-world loop," he said.

Containment and Dual-Purpose Risk

Hassabis described AI as a dual-purpose technology with enormous benefits alongside serious misuse potential. He identified misalignment as a core challenge: ensuring autonomous, agentic systems stick to the goals humans assign them. He said lessons from safety-critical fields like biology and nuclear science should inform AI containment strategies, and called for bringing experts from those areas into the AI safety effort.

The International AI Summit, held during Ireland's EU presidency, drew hundreds of entrepreneurs, lawmakers, and big tech representatives, and focused on harnessing AI for growth.

Sources: Cybernews