METR study shows AI accelerates cyber but shows limited impact on math and AI research

A METR study analyzed AI's contribution across cyber, math, and AI research, finding major acceleration in cyber vulnerabilities (dramatically increased in 2026 vs 2025 for projects like cURL and databases like US NVD), minor acceleration in math (doubled arXiv submissions in some areas, solving problems like the Jacobian conjecture), and no measurable acceleration in AI research across seven problem areas. Separately, researchers built SPADE, a framework for co-evolving environment synthesis and agentic capability through self-play, which significantly improved benchmark scores like AIME and GPQA at the 30B parameter scale.