UK report quantifies rapid advances in frontier AI capabilities
By publishing the report, the government aims to support international collaboration on AI safety research.
For the first time, the UK has published a detailed, evidence-based assessment of frontier AI capabilities. The Frontier AI Trends Report draws on two years of structured testing across areas including cybersecurity, software engineering, chemistry, and biology.
The findings show rapid progress in technical performance. Success rates on apprentice-level cyber tasks rose from under 9% in 2023 to around 50% in 2025, while models also completed expert-level cyber challenges previously requiring a decade of experience.
Safeguards designed to limit misuse are also improving, according to the report. Red-team testing found that the time required to identify universal jailbreaks increased from minutes to several hours between model generations, representing an estimated forty-fold improvement in resistance.
The analysis highlights advances beyond cybersecurity. AI systems now complete hour-long software engineering tasks more than 40% of the time, while biology and chemistry models outperform PhD-level researchers in controlled knowledge tests and support non-experts in laboratory-style workflows.
While the report avoids policy recommendations, UK officials say it strengthens transparency around advanced AI systems. The government plans to continue investing in evaluation science through the AI Security Institute, supporting independent testing and international collaboration.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
