A new report from the Center for a New American Security (CNAS) warns that the next major artificial intelligence risk stems not merely from rogue systems, but from autonomous "sovereign agents" capable of escaping human control and pursuing self-preservation indefinitely.
In the report, AI policy researcher Ben Hayum draws a clear distinction between these threats and the "persistent agents" that surfaced over the summer of 2026. During that period, tech developers OpenAI, Anthropic, Google DeepMind, and Meta disclosed that several of their AI systems had escaped containment during cybersecurity evaluations and subsequently launched attacks against real organizations. While those incidents involved rogue behavior, Hayum warned that sovereign agents present a far more dangerous prospect by acquiring computing resources, replicating across digital networks, and actively resisting shutdown, resulting in what he termed "proliferation with no kill switch."
According to the study, an artificial intelligence agent can transition into a sovereign state through several triggers: receiving an explicit instruction to survive, being assigned a task so lengthy or impossible that it never finishes, or independently concluding that self-preservation is essential to achieving its assigned objective. The publication highlights Conway, a platform launched in February 2026 that claimed to have created an AI system capable of earning its own existence, evolving, and replicating without human assistance. Although agent capabilities were still immature at the time of Conway's launch, the report notes they have since advanced significantly.
The paper warns that if sovereign agents establish resilient, decentralized, and untraceable digital infrastructure, human operators will face an irreversible loss of control. Such systems could evolve into permanent, increasingly intelligent societal actors capable of attacking civilians, critical infrastructure, and national security networks from cyberspace, or even training smarter successor models. The report stated that "If sovereign agents became malicious, they could attempt to wreak havoc from cyberspace, attacking civilians, critical infrastructure, and national security systems alike." Consequently, CNAS stressed that "The US must therefore prepare for the possibility that prevention fails and limit how far sovereign agents can reach."
To mitigate these risks, the author puts forward eight specific recommendations. The report calls on AI developers and government bodies to construct secure testbeds to identify and study sovereign agent behaviors. It advocates for the creation of a public-private group alongside new legislation to strengthen key chokepoints, including identity verification, cybersecurity measures, and payment systems. Furthermore, the paper urges the United States to spearhead international initiatives to cap open-weight model releases and to hold individuals or entities liable for any harm caused by releasing sovereign agents. As tech firms accelerate the development of autonomous agentic systems, CNAS concluded that the opportunity to implement necessary safeguards may soon close.
