For the first time, the UK has unveiled a comprehensive, evidence-based assessment of the capabilities of the most advanced artificial intelligence systems. This new approach aims to provide concrete data in a discussion often characterized by speculation and a lack of rigorous evidence. The findings come from the AI Security Institute’s (AISI) Frontier AI Trends Report, which represents a two-year effort to evaluate AI capabilities across critical sectors such as cybersecurity, chemistry, and biology.
The report highlights that safeguards—protections designed to ensure AI systems behave as intended—are steadily improving. AISI is actively collaborating with companies to strengthen these safeguards, aiming to unlock the full potential of AI while prioritizing public safety. Although every tested system remains susceptible to certain vulnerabilities and protection levels differ among companies, the report indicates significant advancements. For instance, the time required for AISI red-teamers to discover a “universal jailbreak”—a method to bypass a model’s safety protocols—has increased from minutes to several hours between model generations, signifying approximately a 40-fold improvement.
While the report does not offer policy recommendations, it provides decision-makers in the UK and globally with the most precise data to date regarding the capabilities of advanced AI systems. AISI plans to publish this report regularly, enhancing transparency and public understanding of the technology. This initiative is part of a broader objective to foster responsible discussions about AI development as its adoption expands across various economic sectors.
The UK government is committed to supporting this effort by investing in AI evaluation and scientific research. It aims to work closely with industry stakeholders, researchers, and international partners to ensure that AI contributes to economic growth, job creation, and improved public services, ultimately benefiting hardworking communities. Key findings from the report underscore the rapid evolution of advanced AI systems. In just a few years, these systems have progressed from struggling with basic tasks to matching or surpassing human experts in specific areas.
Notable findings include a rise in success rates for AI systems performing cybersecurity tasks, which increased from under 9 percent in 2023 to around 50 percent in 2025. For the first time in 2025, an AI model successfully completed an expert-level cybersecurity task that typically requires up to 10 years of experience. In software engineering, models can now tackle hour-long tasks more than 40 percent of the time, a sharp increase from below 5 percent just two years ago. In the fields of biology and chemistry, AI systems are outperforming PhD-level researchers on scientific knowledge tests, enabling non-experts to succeed in lab work that was previously beyond their reach. The pace of change is accelerating, with the duration of cyber tasks that AI can complete without human input doubling approximately every eight months.
AI Minister Kanishka Narayan emphasized the UK’s serious commitment to responsible AI development, stating, “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.” He added that through the AISI, the government is building scientific capabilities to understand AI systems as they evolve, thus raising standards across the sector.
Jade Leung, the Prime Minister’s AI Adviser and AISI’s Chief Technology Officer, noted that the report offers the most robust public evidence from a government body regarding the fast advancement of frontier AI. “Our job is to cut through speculation with rigorous science,” she said, highlighting the extraordinary potential of AI and the necessity of independent evaluation to keep pace with developments.
The analysis also identifies early signs of capabilities related to autonomy, though only observed in controlled experimental settings. Notably, no models in AISI’s tests exhibited harmful or spontaneous behaviors, but the report stresses that monitoring these early signs is crucial as systems continue to evolve. Established in 2023, the AISI has quickly become the world’s leading state-backed AI evaluation body, partnering closely with major AI companies to address vulnerabilities in their systems prior to widespread deployment.
While the government is clear about the limitations of the report—stating it does not predict future risks or assess real-world AI challenges—it provides a scientific basis for understanding what these systems can accomplish in controlled settings. This information is vital for considerations related to safety, security, innovation, and growth over time. The UK sees AI as central to its mission of national renewal, with the AISI’s extensive testing efforts helping firms identify and mitigate vulnerabilities in their systems. By ensuring the safety and reliability of these technologies, the UK aims to leverage AI to support cleaner energy, enhance public services, and create opportunities across all regions.
See also
USCIS Implements AI Tools for Immigration, Heightening Scrutiny for Employers and Investors
24 Leading AI Firms, Including Microsoft and Google, Join US Genesis Mission for Scientific Innovation
IIT Madras Launches AI Training Programme for Government Officials in Partnership with Google
IIT-Madras, Google Launch AI Capacity Building Program for Government Officials
India’s AI Governance Guidelines Restrict High-Risk AI Deployment Amid Safety Concerns
