METR Finds AI Accelerated Cyber Vulnerability Discovery Sharply, Math Slightly, and AI Research Itself Not at All

Import AI 470 (Jack Clark), on a METR research note·high signalXBlueskyLinkedInCopy link

Import AI 470, published 2026-08-24, walks through a METR research note measuring where LLMs actually moved the needle on discovery. Vulnerability reports accelerated across cURL, OpenSSL, Firefox, Microsoft, the US NVD, and OSV in 2026 versus 2025; math shows arXiv submissions doubling in some areas plus named results including the Jacobian conjecture from Smale's list and Green's Problem 44; and across seven AI optimization targets (CIFAR-10, Hutter compression, Gurobi MIP, MIPLIB, nanoGPT, Stockfish, matrix-multiplication exponent) there is no measurable acceleration. Jack Clark's read is that acceleration arrives as per-skill phase changes, coding in 2025 and cyber in 2026, not as a smooth industry-wide lift.