Deepening AI security research through the UK AI Security Institute (Aisi)

Today, we are announcing an expanded collaboration with the UK AI Security Institute (Aisi) through a new Memorandum of Understanding focused on research and security, to help ensure that artificial intelligence is developed safely and for the benefit of all.

The research partnership with Aisi is an important part of the wider partnership with the UK government in accelerating safe and beneficial progress.

Building on a collaborative basis

AI holds enormous potential to benefit humanity by aiding medicine, accelerating scientific discovery, creating economic prosperity and ameliorating climate change. For these benefits to be realized, we must put safety and responsibility at the heart of development. Testing our models against a wide range of potential risks remains a critical part of our security strategy, and external partnerships are an important part of this work.

That is why we have partnered with UK Aisi since November 2023 to test our most capable models. We are deeply committed to UK Aisi's mission to equip governments, industry and the wider community with a scientific understanding of the potential risks posed by advanced AI and potential solutions and targets.

We are actively working with Aisi to create a robust test of AI models, and our teams have collaborated in security research to move the field forward, including the latest work on hypothetical monitoring: Building on this success, today we are expanding our cooperation in testing to include more, foundational, research in various areas.

Which involves relationships

Under this new research partnership, we are expanding our collaboration to include:

• Sharing access to our proprietary models, data and ideas to accelerate research progress
• Integrated reports and publications to share findings with the research community
• Collaborative security and security and security security that integrate our team's expertise
• Technical discussions to address complex security challenges

Important Research Areas

Our joint research with Aisi focuses on critical areas where Google Dedemind's expertise, diverse teams, and years of pioneering research can help make AI systems safe and secure:

Monitoring AI consulting processes

We will work on strategies to monitor the “thinking” of an AI system, also often called its thinking (COT). This work builds on the previous Google Depinind research and, as well as our recent collaboration on this topic with Aisi, Vulai, Anthropic and other partners. COT monitoring helps us understand how an AI system generates its responses, using translational research.

Understanding the social and emotional implications

We will work together to investigate the effects of social comedy; That is, the power of AI Models to behave in ways that are not compatible with human life, even if they follow the instructions properly. This research will build on the existing work of google depmind which has helped to define this important area of ​​AI in Phezulu.

Evaluating Economic Systems

We will explore the potential impact of AI on economic systems by simulating real-world operations in different locations. Experts will score and validate these activities, after which they will be classified into categories such as difficulty or representation, to help predict long-term market trends.

Working together to reap the benefits of AI

Our collaboration with Aisi is one of the things we aim to do to find the benefits of AI for humanity while addressing the potential risks. Our comprehensive strategy includes privacy research, extensive security training coupled with capacity building, rigorous testing of our models, and the development of better tools and frameworks to understand and mitigate risk.

Strong internal control processes are also essential for safe and responsible AI development, as is working in collaboration with independent external experts who bring fresh ideas and diverse expertise to our work. Google Depmind's Manager of Ethics and Security works across teams to monitor emerging risks, review ethical and security assessments and mitigate technical and policy mitigations. We also collaborate with other external experts such as Apollo research, Vaillis, Dreadnode and more, to carry out extensive testing and testing of our models, including Gemini 3, our most intelligent and secure model to date.

In addition, Google Deepmind is a proud founding member of the Fodeli Model Forum, and the AI ​​partnership, where we focus on ensuring safe and reliable verification of advanced security models.

We hope our increased cooperation with Aisi will allow us to create strong AI security measures for the benefit not only of our systems, but also for everyone who interacts with AI systems.