DeepMind and UK AI Security Institute deepen partnership

Google DeepMind has expanded its collaboration with the UK AI Security Institute, announcing an intensified partnership to advance AI safety and security research. The move, disclosed by both organisations this week, will concentrate on joint work in model evaluation, adversarial testing and the development of shared tools and benchmarks.

The agreement covers multi-institution research projects, exchange of technical expertise and shared access to evaluation frameworks, the organisations said. The partnership is positioned as a research-first effort that aims to produce public datasets, tooling and evaluation methodologies that can be used by the wider safety research community.

What the partnership will do

Under the expanded collaboration, DeepMind and the UK AI Security Institute plan to run coordinated research programmes that target weaknesses in deployed and pre-deployment models. Public statements list three priority areas: improving model evaluation to surface failure modes, developing adversarial testing methods that probe model vulnerabilities, and producing tooling for safer deployment and red-teaming workflows.

The organisations said the work will combine DeepMind’s internal research capacity and model-building experience with the Institute’s focus on policy-relevant security analysis. Planned outputs include reproducible evaluation methods, open-source utilities for adversarial testing and joint papers or technical reports describing findings and recommended practices for model developers.

Researchers from the two organisations will run collaborative experiments and workshops, share non-sensitive evaluation data and coordinate on the design of benchmarks. The partnership is also expected to support training and capacity-building activities for practitioners in the UK and overseas, according to the announcements.

Background and context

The UK AI Security Institute was founded to address security risks arising from advanced AI systems through research, red-teaming and the establishment of measurement standards. DeepMind, a major applied research lab, has previously published work on robustness, alignment and evaluation techniques and has engaged with external institutes on safety research.

This expansion follows an era in which governments, academia and industry have pushed for more systematic evaluation of large models and greater sharing of best practices. The partnership signals continued private-sector interest in coordinated, transparent research to better understand real-world model behaviour and to develop tools that can be widely adopted.

Observers say that joint efforts between a major lab and a specialised security institute can accelerate the creation of community-grade evaluation resources. At the same time, the effectiveness of such collaborations depends on how much material is published openly and whether results are benchmarked against independent baselines.

Why it matters

DeepMind’s collaboration with the UK AI Security Institute brings a major lab’s engineering resources together with a specialised security-focused research body, increasing the chances of producing reusable evaluation tools and documented failure cases. For model developers, regulators and security teams, shared benchmarks and adversarial tooling can make it easier to compare systems and to adopt defensive practices. If the partnership publishes reproducible results and open-source utilities, it could raise the baseline for safety testing across the research ecosystem.