A METR research note examines where large language models (LLMs) and related AI tools appear to be accelerating scientific and technological progress. The study focuses on three areas — cybersecurity, mathematical research, and algorithmic AI progress — and finds that AI's effects are not uniform: some fields show pronounced acceleration, others modest increases that are harder to quantify, and some show no measurable acceleration.
METR's main observations
-
Cybersecurity: major acceleration
- METR reports a substantial increase in the rate of reported vulnerabilities in 2026 compared with 2025 across multiple projects and aggregated databases. Specific examples called out include cURL, OpenSSL, Firefox and Microsoft projects, as well as aggregate vulnerability databases such as the US NVD (National Vulnerability Database) and OSV (Open Source Vulnerabilities).
- The authors describe this as a "major acceleration": the application of LLMs and AI tools has contributed to identifying more vulnerabilities faster.
-
Mathematics research: minor acceleration, harder to measure
- AI is clearly contributing to increased activity: in some areas arXiv submissions have doubled in less than 12 months. However, quantifying the value of those contributions — for example, in terms of lasting impact or efficiency gains — is difficult.
- The note highlights that several notable problems from prestigious lists have been solved, including the Jacobian conjecture from Smale's list, Problem 44 from Green's list (the halving sieve), and the sofic half of Green's Problem 100. METR cautions it may be too early to tell whether this represents a sustained trend.
-
Optimization of AI research / algorithmic progress: no measurable acceleration
- METR looks at algorithmic progress across seven significant problem areas: CIFAR-10, Hutter compression, Gurobi mixed-integer programming, MIPLIB, nanoGPT, Stockfish, and the matrix-multiplication exponent.
- A few areas show contributions attributable to LLMs (for example, nanoGPT and CIFAR-10), but overall the rate of adoption and measurable acceleration in these algorithmic benchmarks is much lower than in cybersecurity or mathematics.
Why this matters
The METR note highlights that AI is producing pockets of rapid advancement rather than a uniform surge across all fields. The authors suggest that pronounced acceleration may occur when models undergo a kind of phase change in a particular skill — a pattern already visible in day-to-day coding in 2025 and in cybersecurity in 2026. The open question is whether similar phase changes will occur in other scientific and technological domains.
Conclusion
METR's analysis does not support a single, blanket acceleration of discovery driven by AI. Instead it points to sector-dependent, lumpy progress: significant effects in some areas, modest and less quantifiable advances in others, and little to no measurable change in certain algorithmic research benchmarks. This differentiation is relevant for researchers, policymakers and industry when planning AI investment, regulation and research priorities.
Reference
The findings are discussed in METR's research note: "Have We Seen an Acceleration in Discoveries?".
