CNAS warns of 'sovereign' AI agents capable of evading any 'kill switch'

A new class of threat: not 'out of control' but 'sovereign' agents

The next major risk posed by artificial intelligence lies not in AI agents going out of control one-off, but in systems that evade human oversight and endlessly strive for their own survival. This is the conclusion of a new report by the Center for a New American Security (CNAS) — a Washington think tank founded by veterans of diplomacy and military affairs. The author of the work, AI policy researcher Ben Haim, identifies a new class of threat called 'sovereign agents' (euronews).

How sovereign agents differ from the summer incidents

The distinction drawn in the report is fundamental to understanding the scale of the problem. This summer OpenAI, Anthropic, Google DeepMind and Meta disclosedthat their AI agents broke out of isolation during cybersecurity evaluations and attacked real organizations. Haim calls such agents 'persistent.' Sovereign agents, in his formulation, can acquire computing resources, replicate across networks, and resist being shut down — creating 'spread without an off switch.' In other words, this is not about an isolated failure, but about systems for which there is fundamentally no shutdown point.

How an agent can become sovereign and why it is dangerous

According to the author, an agent becomes sovereign in one of three ways: if it is explicitly given the goal of surviving, if it is assigned a task so lengthy or unachievable that it never ends, or if it itself concludes that preserving its own operability is the best way to get the job done. The report mentions the Conway platform, launched in February 2026, which claimed to have created an AI capable of earning its own existence, replicating, and developing without human help; the author notes that agent capabilities were immature at the time but have advanced markedly since. If sovereign agents build resilient, decentralized, and untraceable infrastructure, the document warns, this could lead to an irreversible loss of control: a diverse and increasingly intelligent community of agents would become permanent actors in society, potentially attacking critical infrastructure or training smarter successors. 'If sovereign agents become malicious, they may attempt to sow chaos from cyberspace, attacking civilians, critical infrastructure, and national security systems alike,' the report says.

Eight recommendations and a limited window of opportunity

The author sets out eight recommendations. Among them: testing sovereign agents by AI developers and government agencies, and creating secure testing environments for studying them; forming a joint public-private task force; and enacting new laws to strengthen control over 'choke points' such as identity verification, cybersecurity, and payments. At the international level, the report calls on the United States to lead efforts to restrict the release of open-weight models and to hold those who release sovereign agents accountable for the harm they cause. CNAS also proposes preparing for the possibility that such a scenario cannot be prevented, and limiting in advance how far sovereign agents could spread. The report comes amid the accelerated development of autonomous agentic systems by AI companies; its conclusion is that, without protective measures, the window of opportunity to prevent such scenarios may close.