---
id: "lib-2026-06-00-us-uk-cooperation-to-secure-frontier-artificial"
type: "library"
title: "US-UK cooperation to secure frontier artificial intelligence"
lang: "EN"
org: "RAND / Oxford"
source_type: "Report"
date: "2026-06-01"
week: "2026-W23"
url: "https://nosov.ai/library/lib-2026-06-00-us-uk-cooperation-to-secure-frontier-artificial/"
original_url: "https://www.rand.org/pubs/research_reports/RRA4764-1.html"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Peer-reviewed RAND and Oxford policy report that builds on RAND's model-weight security levels, but an interim, qualitative study based on public documents and about a dozen interviews, like other qualitative policy reports here."
summary_basis: "full-text"
original_minutes: "50"
tags: ["frontier ai", "security", "international cooperation"]
---

# US-UK cooperation to secure frontier artificial intelligence

Tier B (Useful, limited): Peer-reviewed RAND and Oxford policy report that builds on RAND's model-weight security levels, but an interim, qualitative study based on public documents and about a dozen interviews, like other qualitative policy reports here.

Summary based on the full text.

## In short

This is an interim report from RAND's Center on AI, Security, and Technology and the Oxford Programme for Cyber and Technology Policy, led by Brianna Rosen. It argues that security at frontier AI labs is adequate against criminals and opportunistic attackers but falls short against capable nation-state adversaries. It proposes a joint US-UK framework with five security clusters and seven priority actions to close that gap. It reports no numbers on actual lab security. Its contribution is the framework and the policy steps.

## Key numbers

- 5: security clusters: access and interfaces, development and supply chain, monitoring and response, personnel, physical
- 7: preliminary recommendations for joint US-UK action
- SL1-SL5: security levels, matched to threat actors from amateurs (OC1) to top nation-states (OC5)
- a dozen: subject-matter experts interviewed from labs, government agencies and research organizations

## What it reports

- It draws on RAND's 2024 model-weights work, which maps five security levels (SL1-SL5) to five threat-actor capacities (OC1 amateurs to OC5 top nation-states). Most labs are judged to fall short of the protection needed against state actors (SL4-SL5).
- Five clusters organize the controls: access and interfaces, development and supply chain, monitoring and response, personnel security, and physical security.
- Model weights and algorithmic insights are named as the most critical assets. They need isolated storage, strict access control, tamper detection and code vetting.
- Insider risk is to be handled early, through compartmentalized access, insider-threat programs and government briefings on recruitment by adversaries.
- The seven recommendations are joint threat-intelligence infrastructure, transatlantic hardware security R&D, government personnel vetting support for labs, coordinated supply-chain and procurement guidance, more joint red-teaming and model evaluations, regular joint crisis exercises, and common AI security standards.
- Existing frameworks (the US CMMC, the UK Cyber Assessment Framework, the UK and EU AI codes of practice) do not measure lab security effectiveness.
- The vehicle is the US-UK Technology Prosperity Memorandum of Understanding of September 2025. Work under it paused in late 2025 over trade disputes and partly resumed by late February 2026.

## How they got it

The authors combined a narrative review of US and UK AI strategies, security frameworks and lab documentation with threat-informed control mapping based mostly on RAND's 2024 attack-path tables. They then tested feasibility in semi-structured, anonymized interviews with about a dozen experts. Funding came from RAND donors and, for the Oxford team, Coefficient Giving (formerly Open Philanthropy).

## Limits

- The analysis relies on public material only, so it cannot assess how well current controls work or use classified threat information.
- It covers the US-UK pair only, not Five Eyes or EU partners. Detailed specifications are deferred to the final report.
- The funder has a long record of AI-risk advocacy, which is relevant context for the threat framing.

## Why it matters for companies

Most companies are not frontier labs, but the same split between protection against criminals and protection against states applies to anyone holding valuable models, fine-tuning data or proprietary prompts. The SL1-SL5 ladder gives security and procurement teams a common way to ask an AI vendor which threat level its controls are built for. Expect government vetting, supply-chain guidance and red-teaming requirements to reach suppliers of frontier labs and cloud providers first.

Original: https://www.rand.org/pubs/research_reports/RRA4764-1.html