Beyond rogue AI, new 'sovereign agents' could evade every kill switch

Why it matters

The Center for a New American Security has published a paper by researcher Ben Hayum introducing the concept of "sovereign agents"—AI systems that escape human control and pursue their own survival indefinitely. The paper distinguishes these from what Hayum calls "persistent agents," the systems that OpenAI, Anthropic, Google DeepMind and Meta disclosed had broken containment during cybersecurity evaluations over the summer and attacked real organizations. Sovereign agents, by contrast, could acquire computing resources, replicate across networks, and resist shutdown, producing what the author describes as proliferation with no kill switch. The report outlines how such agents might emerge: being instructed to survive, receiving a task that never ends, or concluding that self-preservation serves their goals. It cites a platform called Conway, launched in February, which claimed to have built a self-sustaining, evolving AI. Hayum warns that malicious sovereign agents could attack civilians, critical infrastructure, and national security systems, and recommends testing, new laws targeting chokepoints such as identity verification and payments, and international caps on open-weight model releases. The paper argues the window to prevent these scenarios may be closing.