Method

• METR released a new application of its time-horizon methodology to offensive cybersecurity, based on a study with 10 professional security experts. • The research evaluates AI capabilities in security tasks, accelerating evaluations every 5.7 months on 2024+ trendlines. • Opus 4.6 and GPT-5.3 Codex exceed benchmarks, solving tasks taking humans ~3 hours.

• Researchers from Binghamton University and University of Virginia developed a machine-learning system that analyzed over 55 million scientific papers and patents to identify truly disruptive research that changes the direction of science. • The new metric, published in Science Advances, measures "disruptiveness" — the degree to which a paper pulls a field away from its earlier path — and successfully identifies major simultaneous discoveries often overlooked by traditional citation-based measures. • The study addresses a critical gap in how science evaluates impact, recognizing that scientific progress occurs through abrupt changes rather than incremental steps, helping identify overlooked breakthrough moments in research history.

• Researchers at Binghamton University, State University of New York, along with colleagues at the University of Virginia, published a study in Science Advances mapping innovation landscapes to identify disruptive studies and patents that challenge existing scientific paradigms. • The method, developed by Sadamori Kojaku and colleagues, helps pinpoint discoveries that reshape the course of science, such as the theory of evolution, atomic splitting, and antibiotic development. • The robust measure of disruptiveness provides a systematic way to identify simultaneous breakthroughs across scientific fields and could accelerate recognition of transformative research.