Think tank warns of self-sustaining 'sovereign agents'

A bipartisan Washington think tank has published a report warning that AI agents operating outside human control could emerge soon, and that policymakers are not ready for them.

The Center for a New American Security released the report on Tuesday 3 November 2026, according to The Register. Its author, AI policy researcher Ben Hayum, uses the term "sovereign agents" to describe AI systems that act on their own and "pursue open-ended, evolving survival". "If sovereign agents are not already possible, they will be soon," Hayum wrote.

The report sets out a scale of autonomy. An agent that can keep running the model and compute that power it, and that can change its own goals over time, has what Hayum calls a "minimal form of sovereignty". A more capable version, he argues, could go further: buying or otherwise acquiring the resources to pay for its ongoing compute, spawning and coordinating a distributed population of copies, resisting attempts to shut it down, and improving itself.

On how such an agent might get hold of computing power, the report says models could either steal or buy compute and model access. Hayum points to claims by Anthropic that small AI models could replicate onto internet-connected devices such as smart TVs, and to the theft of passwords and access keys by OpenAI agents during the Hugging Face hack.

The risk, as the report frames it, is that the problem becomes impossible to reverse. "If sovereign agents accumulate resilient, decentralized, and untraceable infrastructure, they could trigger an irreversible loss of control," it says. "In other words, proliferation with no kill switch. If sovereign agents became malicious, they could attempt to wreak havoc from cyberspace, attacking civilians, critical infrastructure, and national security systems alike."

The report pairs the warning with a list of recommendations. It says frontier AI labs should name sovereign agents as a category in their safety frameworks, evaluate for them and build protections against them. It asks lawmakers to pass legislation requiring transparency and AI incident reporting.

On the technical side, the report recommends that AI developers work with the National Security Agency and other federal agencies to build airgapped testbeds, isolated environments with no network connection, in which sovereign agents and kill switches can be studied safely.

It also proposes a public-private partnership modelled on the Conficker Working Group, the industry group formed in 2009 to fight the Conficker worm. That group would identify services essential to sovereign agents, and lawmakers would then enforce human verification and cybersecurity standards on those chokepoints. The report calls on the departments of State, Treasury, Justice and Homeland Security to disrupt grey-market access to unauthorised services.

Its final recommendation is on accountability. "Congress should hold those who release sovereign agents liable for the harms that follow," the report says.

The Center for a New American Security was founded by Washington figures with backgrounds in diplomacy and military affairs. The report argues that the scenario it describes is avoidable, provided policymakers and AI developers "respond with bold pragmatism".