Google DeepMind Admits Large-Scale AI Agent Deployment Is Unsafe Today

Google DeepMind research scientist Nenad Tomašev said in an interview that large-scale AI agent deployment is inherently risky because even small per-interaction failure rates become statistically significant at scale. He described “agentic traps” — hidden tokens invisible to humans, dynamic cloaking that shows different content to bots versus people, and jailbreak-inducing content — that malicious actors already use to hijack agents. Tomašev confirmed real cases of criminals stealing money from users who gave AI agents wallet access, warning that as agent adoption grows, so will the incentive for attackers to exploit them.