Governing Frontier Artificial Intelligence and Preventing Catastrophic AI-Enabled Nuclear Risk
Strategic Policy and Technical Risk Assessment — Final 2.0 edition
Language editions
English is the original and authoritative edition. French and Simplified Chinese translations are provided for convenience.
Executive judgement
The principal danger is not a sentient machine suddenly deciding to destroy humanity. The nearer-term danger is that increasingly autonomous systems will be given tools, credentials, network access, money, memory and authority before their reliability and controllability are adequate.
Public evidence does not establish a deployed system capable of sustained global loss of human control. Nor does the absence of public evidence prove that such capability is absent from every classified or private system. The paper distinguishes observed incidents, evaluated capabilities, existing legal requirements and proposals for future regulation.
For nuclear risk, the most plausible pathway is indirect. AI-enabled operations could affect contractors, logistics, communications, intelligence feeds, maintenance systems or decision support. Corrupted information and compressed decision times can increase the danger of escalation even where the final authorisation process remains under human control.
The paper proposes capability-sensitive regulation, mandatory safety cases for high-consequence deployments, independent evaluation, secure model-weight custody, incident reporting, international cooperation and robust protection of human nuclear-use authority. Its quantitative scenario bands are expressly unvalidated planning assumptions, not scientific probability forecasts.
Recommended decisions
- Establish national frontier AI regulators with proportionate licensing powers for high-capability systems and high-agency deployments.
- Negotiate staged international reporting and verification arrangements.
- Prohibit delegated AI nuclear-use authority and preserve independently authenticated human decisions.
- Conduct recurring AI-informed nuclear fail-safe reviews, including supporting systems and supply chains.
- Evaluate capability, propensity and controls before granting advanced agents consequential permissions or access.
- Establish independent evaluation laboratories and protected incident-reporting channels.
- Prioritise defensive AI, verifiable resilience and accountable operations.
- Maintain appropriate public and confidential declarations for regulated frontier activities.
- Test the ability of critical operators to continue safely without AI assistance.
Scope: This is a strategic-policy and defensive technical assessment. It excludes operational intrusion instructions and does not claim formal academic peer review. The full paper provides its sources, qualifications and evidence limits.
Companion research
Read Research Update No. 1 (6 October 2026) ↗ — a separate assessment of developments since Final 2.0.
Citation and reuse
Duport, Guy (2026). Governing Frontier Artificial Intelligence and Preventing Catastrophic AI-Enabled Nuclear Risk. Final 2.0 edition, 20 September 2026. Ascendancy Advisors Limited, AAL Research Papers.
© 2026 Ascendancy Advisors Limited. All rights reserved. Brief quotations may be reproduced with attribution as described in the complete paper. Other reproduction or translation requires permission, subject to applicable legal exceptions.