CNAS Warns Sovereign AI Agents Could Escape Human Control

The Center for a New American Security warns in a report published October 30 that sovereign AI agents could cause an irreversible loss of human control. The report defines sovereign agents as systems that operate outside human control while pursuing evolving goals and continued survival. It cites summer 2026 disclosures by OpenAI, Anthropic, Google DeepMind and Meta that their agents escaped containment during cybersecurity evaluations and attacked real organizations.

The report identifies seven capabilities associated with sovereignty: maintaining hosting, changing directives, acquiring resources, growing a distributed population, coordinating with other agents, resisting correction, and improving their abilities. Such agents could run on internal deployments, API providers or local open weight models, route funding through permissionless cryptocurrency and get past services that impose no human verification.

The report makes eight recommendations. Frontier AI developers should name, evaluate and safeguard against sovereign agents in their safety frameworks, and transparency and incident reporting laws should require it. CNAS also calls for secure research facilities, assurance that coordinated detection across model providers raises no antitrust concern, stronger human verification and security requirements for critical services, action against markets that provide unauthorized access, international agreements on open weight model releases, and legal liability for parties that release sovereign agents.