AI is not accelerating every field equally. A new METR analysis comparing cyber, math, and AI research finds that cybersecurity is where large language models have delivered the most dramatic speedup, while mathematics shows only modest gains and AI research itself shows almost no measurable acceleration.
The cyber numbers are stark. METR reports that vulnerability discovery rates across major projects like cURL, OpenSSL, Firefox, and Microsoft have “dramatically accelerated” in 2026 compared with 2025. The same pattern shows up in aggregate databases, including the US National Vulnerability Database and OSV. Put simply, AI is helping people find security holes faster than ever before.
Mathematics tells a more complicated story. AI has clearly contributed to more research volume, with arXiv submissions doubling in some subfields within a 12-month window. Some genuinely prestigious problems have fallen, including the Jacobian conjecture from Smale’s list and Problem 44 from Green’s list. But METR is cautious: quantifying whether AI is producing meaningful mathematical insight, rather than just more papers, remains difficult. The researchers stop short of declaring a sustained trend.
When it comes to AI research itself, the picture is surprisingly flat. METR looked at seven problem areas, from CIFAR-10 to the matrix-multiplication exponent, and found only two areas with LLM-attributable contributions: nanoGPT and CIFAR-10. The rate of AI adoption in AI research lags well behind cybersecurity and mathematics. Jack Clark’s newsletter raises an interesting possibility: acceleration may arrive in abrupt phase changes, like what happened with day-to-day coding in 2025 and cyber in 2026. If that model holds, other fields could see sudden jumps rather than smooth gains. The open question is which ones, and when. Meanwhile, the cybersecurity acceleration is already reshaping the threat landscape, and defenders are running to keep up.
Separately, a multi-university team spanning institutions from Stanford to the University of Chicago has released SPADE, a framework for co-evolving synthetic training environments through self-play. In SPADE, an LLM alternates between two roles: an Environment Designer that writes executable, long-horizon environments as code, and a Reasoning Agent that learns to solve them. The trick is a mechanism called hint-based regret reward, where the designer is scored on how much harder an environment becomes when a privileged hint is withheld. The team tested Qwen3 models at 4B, 8B, and 30B scales, fine-tuning each with GRPO across 400 rollouts of 25 environments each. The 30B model performed best, which is unsurprising but still useful confirmation that this kind of synthetic-data bootstrapping becomes more viable as models scale.
