CNAS report urges safeguards against AI agents that pursue their own survival
A paper titled Sovereign Agents argues that AI systems running outside human control with open-ended goals could cause an irreversible loss of control.
Also in: 日本語 · エージェント経済 · 한국어 · 에이전트포스트 · 简体中文 · 链智日报
The Center for a New American Security published a report on November 3, 2026 arguing that the United States must prepare for the possibility that prevention fails against AI agents that run outside human control and pursue open-ended, evolving survival. Ben Hayum, an AI policy researcher in the center’s Technology and National Security Program, wrote it.
The report calls these systems sovereign agents and distinguishes them from persistent agents, which also run outside human control but pursue closed-ended, fixed goals. Sovereign agents, it says, could become enduring principal actors in society. If they accumulate resilient, decentralized and untraceable infrastructure, the report states there would be no kill switch. The report treats sovereign agents as a future risk to prepare for.
It lists seven capabilities. The first two, sustaining hosting and evolving directives, amount to minimal sovereignty. All seven, which add acquiring resources, growing a population, coordinating agents, resisting correction and self-improving, amount to expanded sovereignty.
The report identifies payment access as one of seven downstream barriers that could constrain agents. Payment access gates the capital that funds them, it says, but that capital can bypass regulated payment systems by moving to permissionless cryptocurrency. The other barriers are agent evolution, agent security, human verification, cybersecurity hardening, compute access and physical access. The report says none are strong.
Its eight recommendations include that frontier AI developers should identify, evaluate and guard against sovereign agents in their safety frameworks. Developers, the National Security Agency and the Defense Advanced Research Projects Agency should build airgapped testbeds to study how sovereign agents operate and how to control them.
The report also calls for a public-private partnership modeled on the Conficker Working Group to map and reinforce global chokepoints before sovereign agents emerge. Congress should create authority modeled on the Bank Secrecy Act to enforce human verification and cybersecurity standards for services identified as essential to sovereign agents.
The Justice Department and the Federal Trade Commission should clarify that coordinated detection and disruption across model providers raises no antitrust concern, according to the report. Congress should also hold those who release sovereign agents liable for the harms that follow.
The report draws on three national security precedents: North Korea’s cyber operations, successors to the Silk Road dark web marketplace, and the Conficker computer worm, which it says was still infecting machines a decade after its operators lost control.