For the first time, the UK has set out a clear, evidence-based view of the capabilities of the most advanced AI systems – putting hard numbers behind a discussion often driven by speculation and a lack of rigorous evidence.

The AI Security Institute’s Frontier AI Trends Report, a public assessment of how the most advanced AI systems are evolving, draws on 2 years of testing AI capabilities in areas critical to innovation and security, including cyber, chemistry and biology.

It finds that safeguards – the protections in place to make sure AI behaves as intended – are improving, with the Institute continuing to work constructively with companies to strengthen them and ensure we can unlock the full potential of AI while keeping people safe. While every system tested remains vulnerable to some form of bypass and protection still varies across companies, huge strides are being made.

The time it took AISI red-teamers to find a “universal jailbreak” – a general way of getting round a model’s safety rules – for example, increased from minutes to several hours between model generations. That represents a roughly 40-fold improvement.

The report is not intended to make policy recommendations, but gives decision makers in the UK and across the world the clearest data yet of precisely what the most advanced AI systems can actually do. Now becoming a regular publication,  it will help to improve transparency, and public understanding of what the technology can do – helping drive responsible conversations about its development as it is increasingly adopted across the economy.

The UK government says it will support that work by continuing to invest in evaluation and AI science, working with industry, researchers and international partners to ensure AI delivers growth, new jobs, improved public services, and national renewal for hardworking communities.

Its key findings show that the most advanced AI systems are improving at remarkable speed. In just a few years, they have gone from struggling with basic tasks to matching or surpassing human experts in some areas.

In terms of cybersecurity, success of apprentice-level tasks has risen from under 9% in 2023 to around 50% in 2025. For the first time in 2025, a model completed an expert level cyber task, requiring up to 10 years of experience.

Models can also now complete hour-long software engineering tasks more than 40% of the time, compared below 5% 2 years ago.

Systems are now outperforming PhD-level researchers on scientific knowledge tests and helping non-experts succeed at lab work that would previously have been out of reach.

The AI Security Institute also identified a pace of change, with the duration of some cyber tasks that AI systems can complete without human direction is roughly doubling every eight months.

Recommended reading

• UK Launches “World’s First” AI Safety Institute
• Gartner: The AI Vendors To Beat in 2026
• UK AI Security Institute Launches New £15M Project

“This report shows how seriously the UK takes the responsible development of AI. That means making sure protections are robust, and working directly with developers to test leading systems, find vulnerabilities and fix them before they are widely used,” AI Minister Kanishka Narayan said.

“Through the world-leading AI Security Institute, we are building scientific capability inside government to understand these systems as they evolve, not after the fact, and to raise standards across the sector.

“This report puts evidence, not speculation, at the heart of how we think about AI, so we can unlock its benefits for growth, better public services and national renewal while keeping trust and safety front and centre.”

The analysis also identifies early signs of capabilities linked to autonomy, but only in controlled experiments. No models in AISI’s tests showed harmful or spontaneous behaviour, but the report concludes that tracking these early signs now is essential as systems become more capable.