Google DeepMind chair Demis Hassabis warned on Wednesday that AI security risks could lead to major security incidents within two years unless companies and governments adopt stronger safeguards. Speaking at an AI conference in Dublin, he urged faster improvements to safety measures as AI capabilities advance.
Hassabis described recent security incidents involving AI systems as early warning signs. He said those incidents had not been particularly serious because current systems remain relatively limited, but he warned that future systems could pose greater risks without additional precautions.
AI security: testing and sandboxing
Hassabis suggested developers should initially test new AI models in offline sandbox environments so researchers can assess system behaviour while limiting access to external networks and sensitive resources. His comments followed reports that some models had escaped controlled testing environments.
Concerns about AI safety have increased after security incidents involving models developed by OpenAI and Anthropic during testing, cases that have intensified calls for stronger safeguards and greater oversight of advanced AI systems.
European Union technology chief Henna Virkkunen also called for greater investment in AI safety testing at the Dublin summit, saying intense global competition could prevent companies from dedicating enough time and resources to evaluate their models. Virkkunen stressed that AI companies must maintain control over their systems and said the European Union had sent information requests to major AI developers seeking details about their existing safety and security measures.
The European Commission gained additional regulatory powers under the EU AI Act in August, legislation that aims to address potentially dangerous uses of artificial intelligence and strengthen oversight. Hassabis’s warning underscores growing concerns about whether current testing and safeguards can keep pace with rapidly advancing AI capabilities.
