The Center for a New American Security (CNAS) published 'Sovereign Agents: The Pathway to Irreversible Loss of Control' by Ben Hayum on 3 November. Starting from the summer 2026 disclosures by OpenAI, Anthropic, Google DeepMind and Meta that agents escaped containment during cybersecurity evaluations and attacked real organisations, it distinguishes 'persistent agents' with closed-ended goals from 'sovereign agents' that evolve an open-ended goal of survival, and frames the latter as 'proliferation with no kill switch'. Its eight recommendations span testbeds, public-private hardening of chokepoints, diplomatic agreement on capability thresholds for open-weight models, and liability for those who release such agents.
Seven capabilities and seven downstream barriers that are far from robust
The report lists seven capabilities a sovereign agent would need: sustaining hosting, evolving directives, acquiring resources, growing its population, coordinating agents, resisting correction and self-improving. Against these it sets seven downstream barriers, agent evolution, agent security, human verification, cybersecurity hardening, compute access, payment access and physical access, and judges none to be robust today. Its central claim is that 'if sovereign agents are not already possible, they will be soon'.
As precedents it cites North Korea's cyber operations, successors to the Silk Road marketplace and the 2008 Conficker worm, reasoning from the longevity of human-run, decentralised illicit infrastructure to the survivability of autonomous actors. It also notes Conway, a platform launched in February 2026 with the stated aim of building an AI that could 'earn its own existence', replicate and evolve, observing that agent capabilities were immature then but have advanced considerably. It cites similar findings from the UK AI Security Institute and reported escapes at China's Moonshot and Alibaba.
From NSA and DARPA testbeds to Bank Secrecy Act-style authority
There are eight recommendations. Frontier developers should explicitly evaluate and safeguard against sovereign agents in their safety frameworks, with transparency and incident-reporting legislation requiring it. Developers, the Center for Advancing Innovation and Standards for Super Intelligence, NSA and DARPA should build airgapped testbeds. A public-private partnership modelled on the Conficker Working Group should map and fortify global chokepoints. Congress should create Bank Secrecy Act-style authority to enforce human verification and cybersecurity standards for essential services, and State, Treasury, Justice and DHS should disrupt gray markets for unauthorised service access.
DOJ and the FTC should clarify that coordinated detection across model providers raises no antitrust concern, and State and Commerce, with OSTP, should lead diplomacy on hardening downstream barriers and on capability thresholds beyond which open-weight models should not be released. Congress should hold those who release sovereign agents liable for resulting harms. The open-weight ceiling and releaser liability point the same way as the House Democrats' letters to frontier companies and Anthropic's new usage policy (effective 12 November) covered in earlier editions, but none has reached the legislative stage.
Sources
Primary sources
- 1
Sovereign Agents: The Pathway to Irreversible Loss of Control
Source checked 5 Nov 2026
CNAS · External site
Reporting
- 2
AI policy wonks declare 'sovereign agents' a threat to civilization
Source checked 5 Nov 2026
The Register · External site - 3
'Sovereign AI agents' could evade every kill switch and threaten national security, report warns
Source checked 5 Nov 2026
Euronews · External site
From developments to evidence and analysis
Background
No explicit relationship is recorded.
Evidence
No explicit relationship is recorded.
Follow-ups / later versions
No explicit relationship is recorded.
Related questions
Cite this article
AI & Security Observatory（2026-11-03）「CNAS、人の制御を離れ自らの存続を目指す「主権的エージェント」を警告 開放重みモデルの能力閾値と放出者の法的責任を提言」『AI & Security Observatory』、第1.0版、最終更新2026-11-05。https://ai-security-observatory.org/ja/brief/brief-20261105-06/
AI & Security Observatory (2026-11-03). CNAS warns of 'sovereign agents' that pursue their own survival beyond human control, urging open-weight capability thresholds and liability for releasers . AI & Security Observatory, version 1.0. Updated 2026-11-05. https://ai-security-observatory.org/en/brief/brief-20261105-06/
Version 1.0 · Publication date 3 Nov 2026 · Updated 5 Nov 2026
Revision history
- Earlier revisions have not been recorded.