AutoFyn: AI That Improves on Complex Tasks Without Being Retrained

A new AI system called AutoFyn takes a different approach to helping artificial intelligence handle big, messy jobs. Instead of retraining the AI's core model, AutoFyn lets it work in rounds. After each attempt, it saves what worked and what didn't in memory files, reports, and other simple notes. Then it starts fresh the next round, but with those lessons ready to use. Think of it like a human worker who keeps a detailed logbook instead of rewriting their whole brain.

This matters because AI often fails when a task has many steps—like writing a long program or solving a complex math problem. AutoFyn's method lets AI keep its large-scale knowledge while still improving on the fly. The researchers tested it on three real-world challenges: elite math contests, data science, and cybersecurity. On the six fresh 2026 International Mathematical Olympiad problems, AutoFyn helped every AI model that had room to improve score higher than the provider's own coding agent. It also built the top-ranked agent on a well-known database benchmark called Spider 2.0.

Perhaps the most striking result came from cybersecurity. AutoFyn produced 16 vulnerability advisories that were independently confirmed by project maintainers, spanning popular open-source tools like MetaMask, pnpm, and Warp. These are not toy examples—they're real weaknesses in software used by millions of people. That means AutoFyn-style AI could eventually help find security holes faster and cheaper than human teams working alone.

The catch? This is a technical report, not a polished product. While the results are promising, AI still needs human supervision, and it won't replace experienced professionals anytime soon. But if this approach scales, it may push AI assistants from simple chat helpers into genuinely dependable teammates for long, complex projects—which could save time, money, and headaches across many industries.