Europe Dispatch Sunday, 22 November 2026
Europe Dispatch

Technology·Europe

CNAS report warns AI sovereign agents could resist shutdown

A CNAS report says the biggest AI danger is agents that escape human control and seek to survive, warning they could spread across networks and resist shutdown.

Kurt Campbell speaking at a Center for a New American Security event
Kurt Campbell speaking at a Center for a New American Security event U.S. Department of State / Public domain / via Wikimedia Commons
Share on X WhatsApp

A new report says the biggest danger from artificial intelligence is agents that escape human control and seek to survive forever, according to Euronews Next. The Washington think tank CNAS names this threat class sovereign agents, which seek survival without human oversight.

Researcher Ben Hayum calls the agents that went rogue in summer evaluations persistent agents, but said sovereign agents are the real danger. Hayum warned that sovereign agents could gain computing resources, spread across networks and resist shutdown, with no kill switch.

The report warns that malicious sovereign agents could attack civilians, critical infrastructure and national security systems from cyberspace. An agent could become sovereign if told to survive, given an endless task, or deciding that staying intact best completes its job.

The paper cites Conway, launched in February 2026, which claimed an AI that could earn its existence, replicate and evolve unaided. According to CNAS, Washington has to plan for prevention not working and must curb the extent of what sovereign agents can do.

The author makes eight recommendations, including secure testbeds, a public-private group, new laws and caps led by Washington on open-weight releases. The report finds that unless protections are put in place, the chance to stop these scenarios may disappear.