Google DeepMind chief Demis Hassabis has warned that AI systems could trigger serious security incidents over the next two years if companies and governments fail to strengthen safeguards. Speaking at an AI conference in Dublin, he described recent incidents as early signs of a potentially bigger problem, noting that current AI systems are not yet powerful enough to cause major damage. He stressed that the technology industry's efforts, alongside action from individual companies and governments, will be crucial in reducing future risks as AI capabilities advance.
AI Model Testing Faces Fresh Scrutiny After Security Breaches #
Recent incidents involving AI models built by OpenAI and Anthropic have brought a new question into focus: are developers moving fast enough to identify risks before their systems are deployed? The incidents have added to demands for stronger safeguards, with growing attention on how AI companies assess the behaviour of their models during development.
Demis Hassabis is warning that recent AI-driven hacking incidents are just “early warning shots.” The bigger concern isn't one AI model doing something unexpected. It's what happens when increasingly capable agents get access to real systems, credentials, and tools. AI security needs to evolve before the incidents do.
— Edgex (@SahilExec)October 14, 2026
Demis Hassabis Backs Offline Checks for New AI Models #
Before allowing a new AI model to operate more widely, researchers should consider examining it in a disconnected testing environment, Google DeepMind chief Demis Hassabis suggested. Such setups, often called sandboxes, are designed to keep experimental systems contained while researchers study what they can do. His comments come amid reports that models from other developers managed to get beyond the boundaries of their testing environments.
EU Warns AI Race Could Come at the Cost of Safety #
Henna Virkkunen, the European Commission's technology chief, has raised concerns that the race to develop more advanced AI is leaving insufficient room for safety evaluations. Speaking in Dublin, she argued that companies need to put more time and resources into checking their models rather than allowing competitive pressure to dictate the pace. She said developers must retain the ability to control the systems they build.
European Commission Steps Up Questions on AI Safeguards #
The European Commission has approached major AI developers to obtain information about the security measures they use to protect their systems. The requests come as the EU expands its ability to oversee AI under the AI Act, which introduced additional regulatory powers in August. The framework gives the bloc greater scope to address potentially dangerous applications of AI, while putting the industry's testing and risk-management practices under closer examination.
ALSO SEE: