The UK’s AI Security Institute released its first Frontier AI Trends Report on December 18, 2025, providing the most detailed evidence-based assessment of advanced AI capabilities to date. Drawing on two years of rigorous testing across more than 30 unnamed cutting-edge models, the report reveals that the performance of these systems is doubling in some critical areas every eight months. In the cyber domain, AI success on apprentice-level tasks surged from 10% to 50% in just one year, with some models now successfully completing expert-level tasks that typically require a decade of human experience.
The research highlights a significant jump in scientific proficiency, with AI systems now outperforming PhD-level researchers in chemistry and biology knowledge tests. These models are increasingly capable of generating accurate laboratory protocols and providing troubleshooting support for wet-lab experiments that is 90 percent more effective than human experts. The report also identifies a steep rise in autonomous “agentic” capabilities, where systems can now complete multi-step software engineering tasks lasting over an hour without human intervention, compared to a success rate of less than 5 percent just two years ago.
On the safety front, the Institute found that while methods used to bypass a model’s safety rules known as “universal jailbreaks”, now take hours rather than minutes to achieve, every system tested remains vulnerable to some form of bypass. While two cutting-edge models showed a 60% success rate in self-replication tests, no models demonstrated spontaneous attempts to replicate or hide their capabilities. Societal data within the report also indicates that one in three UK citizens now uses AI for emotional support or social interaction, highlighting a rapid shift in how the technology is being integrated into daily human life.
