@ai_and_law · Post #732 · 26.12.2025 г., 08:04
🇬🇧UK Publishes First Evidence-Based Assessment of Frontier AI Capabilities The UK AI Security Institute released its inaugural "Frontier AI Trends Report", presenting a public, data-driven assessment of how the most advanced AI systems are evolving. Based on two years of testing across cyber security, software engineering, biology, and chemistry, the report provides quantified evidence on AI capabilities, replacing speculation with measurable benchmarks. The findings show rapid capability growth. In cyber security, success on apprentice-level tasks rose from under 9% in 2023 to about 50% in 2025, and for the first time a model completed an expert-level task requiring up to 10 years of experience. In software engineering, models now complete hour-long tasks over 40% of the time, up from below 5% two years ago. In biology and chemistry, systems outperform PhD-level researchers on knowledge tests and enable non-experts to conduct advanced lab work. Safeguards are improving but remain imperfect. The time needed to discover a “universal jailbreak” increased from minutes to several hours between model generations, around a 40-fold improvement, though all tested systems remain vulnerable to some bypasses. The report makes no policy recommendations, but aims to improve transparency and inform regulators and policymakers globally about what frontier AI systems can actually do. #AIRegulation#AISafety#UKAI#FrontierAI#AIGovernance#TechPolicy