Has AI accelerated discoveries? METR seeks inflection in public data

By tracking slope breaks in several time series, METR finds a spectacular acceleration in cybersecurity, plausible in mathematics, and unobtainable in algorithmics.

Published on August 14, 2026, this METR note asks a simple but rarely methodically addressed question: if AI systems really do accelerate research, this acceleration should be visible in public data. The authors therefore looked for slope changes in time series covering cybersecurity, mathematics and algorithmic optimization, testing early 2026 as a possible inflection point.

The clearest signal comes from software security. The number of CVEs identified on cURL increases from 9 in 2025 to 36 in mid-2026; Firefox climbs from 210 to 342; OpenSSL from 6 to 39. In mathematics, the picture is more ambiguous: arXiv submissions in combinatorics have doubled, from 377 to 743 per month, and three open problems of classic lists were solved in 2026 with the help of models — but the share attributable to AI remains difficult to isolate.

On the other hand, on the algorithmic optimization benchmarks publicly monitored — nanoGPT, CIFAR-10, Stockfish — no acceleration is detectable. The authors themselves describe this contrast as enigmatic, even though announcements of results obtained with little inference calculation were circulating at the beginning of 2026. The conclusion is deliberately heterogeneous: depending on the domain observed, acceleration is massive, probable or absent.