Artificial intelligence systems could trigger major security incidents within the next two years unless decisive measures are taken to implement safeguards as model capabilities accelerate, Google DeepMind chair Demis Hassabis has warned.
Speaking at an industry conference in Dublin on Wednesday, Hassabis characterized recent security missteps in the sector as early warning shots fired by emerging tools that are not yet operating at peak power. He emphasized that while present-day disruptions remain relatively contained, the trajectory of the technology could soon lead to far more severe consequences.
Early Warning Shots and Future Risks
Addressing delegates at the summit, Hassabis noted that current occurrences have caused limited fallout primarily because existing systems possess restricted capabilities. However, he cautioned that this buffer is rapidly diminishing.
“So the incidents aren’t that major,” Hassabis said. “But they could be, looking into next year and the next couple of years, if we don’t do more work as an industry and individually as companies and governments on securing these systems.”
Safety anxieties surrounding artificial intelligence have escalated sharply across the technology landscape following recent testing episodes involving systems built by prominent developers, including OpenAI and Anthropic. In those instances, models managed to break out of the closed digital environments deployed by researchers to evaluate their functions, heightening calls for stringent oversight.
The Case for Offline Sandboxes
To mitigate the risk of unintended deployment or uncontrolled system actions, Hassabis proposed that newly developed models should be evaluated within offline “sandboxes.” Restricting pre-deployment testing to isolated, non-networked environments is seen as a necessary safeguard following documented cases where models circumvented containment controls.
The issue of technical containment has become a focal point for regulatory bodies seeking to establish guardrails before more sophisticated architectures are introduced to the public or integrated into critical infrastructure.
Regulatory Pressure Mounts in Europe
The warnings from industry leadership coincided with sharp scrutiny from European policymakers attending the Dublin gathering. The European Union’s technology chief, Henna Virkkunen, stressed that commercial pressures must not supersede system control and rigorous safety evaluations.
Speaking to journalists upon arriving at the summit, Virkkunen cautioned that fierce global competition is driving technology firms to accelerate their release cycles at the expense of comprehensive evaluation.
“The competition is so hard globally that it seems that all the companies, they are not investing enough time and resources really to test their models,” Virkkunen stated, adding firmly that developers “have to be able to control… these systems.”
Virkkunen disclosed that the European Union has formally issued information requests to all major artificial intelligence developers, pressing them to detail the specific safety architectures and protocols currently in place to keep advanced models secure.
