As we close out 2025, I’ve found myself reflecting on what an extraordinary year it has been for the AI Security Institute (AISI). When I joined only a few months ago, I was instantly struck by what an incredibly unique team of talent has assembled, experts with direct experience of the frontiers of AI, drawn from around the world. People who have answered the call to help navigate AI’s challenges safely and securely for decades to come.
This year, that mission sharpened and accelerated in ways that make me incredibly proud of this organisation and the people in it. We’ve shifted to a clearer, more focused remit, reflected in our new name: the UK AI Security Institute. And across technical science, government preparedness, international partnerships, and the tools we’ve put in the hands of the wider ecosystem, AISI has stepped up in a way few would have thought possible two years ago.
Everything we have achieved this year points in one direction: rigorous science, turned into practical action, so the UK can stay safe while realising AI’s enormous potential.
Below are ten achievements that capture the scale and direction of our work this year.
1. We delivered the UK government’s first Frontier AI Trends Report
Last week, we released our first Frontier AI Trends Report, providing a clear, evidence‑based picture of how frontier AI capabilities are evolving, based on two years of AISI’s testing.
This is a major milestone for the UK and for AISI. For the first time, we’ve put hard numbers behind what the most advanced AI systems can actually do. The trends we highlight – rapid capability growth across cyber, biology, chemistry and software engineering, alongside improving safeguards – underline why having this capacity inside the UK government matters. It gives the public, policymakers and our partners a clear view of AI progress that is grounded in data rather than speculation.
2. We tested more frontier AI systems than ever before
Our technical team has now tested more than 30 of the world’s most advanced models. Developers continue to work with us because our evaluations are rigorous, reproducible and grounded in real‑world risk. This year alone we stress‑tested agentic behaviour, and deepened our cyber, chem‑bio and alignment assessment suites. We pioneered new ways to assess novel risks such as developing a purpose-built benchmark to track early signs of self-replication, and experimenting with methods to detect sandbagging, where AI models downplay their own capabilities during testing. We also published our first paper in Science; a large-scale study with over 76,000 participants exploring the levers of AI-enabled persuasion.
3. We advanced the science of evaluation
AISI’s role isn’t just to run tests – it’s to make evaluation itself more scientific and reliable.
This year we introduced new frameworks such as Bayesian GLMs for evaluator reliability and transcript‑level analysis to better understand agent behaviour. We’ve contributed to work analysing the robustness of existing benchmarks as well as best practices for rigorous agentic testing.
4. We uncovered and helped fix critical vulnerabilities
We’ve collaborated with AI developers to identify safeguard vulnerabilities and drive concrete mitigations. Our end‑to‑end biosecurity red‑teaming with OpenAI and Anthropic revealed dozens of vulnerabilities including new universal jailbreak paths. Agents revealed dozens of vulnerabilities including new universal jailbreak paths. We also conducted the largest study of backdoor data poisoning to date with Anthropic, demonstrating how even tiny corruptions can propagate through training, and an agent red‑team with Grey Swan that identified 62,000 vulnerabilities across sectors.
5. We’ve expanded the world’s most widely used government-backed evaluation tools
Inspect, InspectSandbox, InspectCyber, and our latest release, ControlArena, are now being used by governments, companies, and academics around the world. These open tools lower the barrier to high‑quality evaluation and make safety science accessible at scale.
6. We invested in the broader ecosystem
AI security cannot be solved by one organisation. This year alone we launched or expanded:
- The £15m Alignment Project – now one of the largest global alignment research efforts.
- The £8m Systemic Safety Grants – strengthening resilience across society
- The £5m Challenge Fund – unlocking rapid advances on urgent questions
These programmes attracted thousands of proposals and are fuelling cross‑disciplinary progress.
7. We prepared government for plausible futures
AISI ran multiple exercises and briefings with national security partners and across government to help teams plan for emerging capabilities and coordinate faster during crises. These exercises ensure that the UK is not caught off guard when the frontier shifts. As MI5’s Director‑General put it, we must “think deeply, today, about what defending the realm looks like in the years ahead.” This year, we did exactly that.
8. We’ve advanced international collaboration
This year marked a step forward in how countries work together on the science of AI measurement. The international network we helped launch two years ago has now evolved into the International Network for Advanced AI Measurement, Evaluation and Science, a shift that reflects a shared commitment to rigorous, evidence‑driven approaches to understanding advanced AI.
9. We expanded our collaboration with leading developers
Last month we launched a new research partnership with Google DeepMind, enabling shared data access, joint publications, and deeper collaboration on alignment‑relevant questions - including chain‑of‑thought monitoring, socio‑affective alignment, and AI’s impact on economic activity. We also deepened collaborations with Anthropic, OpenAI, and Cohere as part of broader MOUs between the UK government - which all include commitments to work with AISI.
10. We strengthened AISI itself – our mission, our culture, and our leadership
This year we took major steps to clarify our strategy, refine our culture, and build a world‑class team with the urgency and ambition our mission demands. We focused on inclusion, high performance, and building a multidisciplinary environment where talented people can do their best work. And we did it while growing our extraordinary team of researchers, engineers, policy specialists and operators. The energy and camaraderie I’ve seen across AISI makes me confident we can meet the scale of the challenge before us.
Looking ahead
AI is advancing extraordinarily fast. Our job is to stay ahead of this curve – through science, partnerships, and the evidence governments need to act with proportion and confidence. The work is urgent. But everything I’ve seen since joining tells me that AISI has the mission clarity, and the people to meet it.
Thank you to everyone across AISI and to our partners across government, industry, and academia who have supported and challenged us this year. Here’s to another year of science, impact and progress on one of the most important challenges of our time
