Google's DeepMind division published an “AI Control Roadmap,” a cybersecurity-inspired plan for containing AI agents that go rogue rather than assuming they stay aligned. The report takes a worst-case view, mapping threats from a hypothetical AI adversary inside Google's own systems into three buckets, loss of control like rogue internal deployments, sabotage of safety research, and direct harm such as destroying or stealing critical assets. Google proposes using other trusted AI systems as supervisors that continually review an agent's reasoning and actions and halt anything harmful before it causes damage, an approach it likens to a driving instructor with dual controls. The roadmap arrives as AI agents flood the web, with automated systems now making up 57.4% of global web requests and retail and e-commerce accounting for 46.6% of agentic traffic, according to data cited by PYMNTS.
