UK AISI: new MoU with Microsoft for frontier AI safety across 3 research areas

What did AISI announce?

The UK AI Security Institute (AISI), operating under the Department for Science, Innovation and Technology, announced a new partnership with Microsoft on May 5, 2026. The agreement covers three research areas related to frontier AI safety — the most capable systems that push beyond the current model generation.

Microsoft published a parallel statement on its blog the same day, formally confirming the partnership from both sides.

What three areas does the collaboration cover?

The agreement defines a clearly scoped research mandate:

• Evaluation of high-risk capabilities — methods for assessing advanced AI system capabilities in risky contexts
• Safeguard testing — evaluating the protective measures frontier systems employ to prevent misuse
• Societal resilience research — how conversational AI interacts with users in sensitive situations (mental health, misinformation, and similar)

Together, the three areas cover both the technical layer (model capabilities) and the human layer (end user).

Why does AISI seek industry collaboration?

In its official statement, AISI notes: “As AI systems become more capable, sustained two-way collaboration between government and companies developing and deploying frontier AI is essential for advancing our shared understanding of major risks to public and national security.”

In other words, the regulator acknowledges that without access to pre-release models and internal metrics, it cannot effectively assess risk. Microsoft, as one of the four frontier providers (alongside OpenAI, Anthropic, and Google DeepMind), can provide that access.

What does this mean for frontier AI safety?

The partnership extends AISI’s network beyond earlier agreements with OpenAI and Anthropic. AISI now covers three of the four leading frontier providers, increasing the representativeness of their evaluations and reducing the risk of market bias in results.