At the Dublin AI summit on October 14, 2026, Google DeepMind president Demis Hassabis warned that AI agents could cause serious incidents within one or two years if companies and governments failed to strengthen safety measures, according to AFP’s account published by ABC Color. Hassabis said the loss-of-control incidents he had seen so far were warning signs, though in his view they had not been serious.
He recommended testing new models first in offline sandboxes: isolated environments where a model can be assessed without internet access. That is one proposed testing measure. Google DeepMind’s own roadmap describes a broader set of controls for its internal agents.
How Google DeepMind describes agent controls
In a June 18, 2026 post, Google DeepMind outlined its AI Control Roadmap, a layered approach that treats internal agents as potential insider threats. It combines traditional security measures with model alignment, graduated permissions tied to verified behavior, trusted AI supervisors and monitoring intended to detect misaligned actions or attempts to evade oversight.
The roadmap also matches intervention to risk. For low-risk actions that can be reversed, Google DeepMind describes reviewing activity later and addressing problems then. For particularly high-risk actions, such as major cyberattacks, it describes real-time prevention. The company says it assesses monitoring through three measures: coverage, or how much traffic is monitored; recall, or how much misaligned behavior is caught; and time to respond.
That design assumes a model may not always follow its intended goals. Monitoring and permission limits are meant to help catch problems and restrict what an agent can do; they are safeguards, not a guarantee that every failure will be prevented.
What Google DeepMind says its monitoring has covered
Google DeepMind says an internal prototype analyzed one million coding-agent tasks and helped inform a live monitor for its Gemini Spark agent. The company also says most flagged events in its data arose from misinterpretation or overeagerness rather than adversarial intent. That distinction matters: an action can be unsafe without the agent trying to cause harm.
Common security risks for AI agents
An agent typically combines a model with tools and a repeated action loop. That can let it do more than generate a response, but it also creates more ways for instructions or behavior to go wrong. In an English-language explainer, IBM Technology’s Jeff Crume discusses risks including goal hijacking, rogue agents and memory poisoning. These are general security examples, separate from Hassabis’s summit warning and Google DeepMind’s internal roadmap.
