AI researchers say self-improvement is here: 80% of Claude's code now written by Claude

The idea that AI systems will eventually build better versions of themselves used to sound like science fiction. After interviewing 25 researchers from OpenAI, Anthropic, Google DeepMind, Meta, and US universities, IAPS fellow Severin Field says that framing no longer holds. Twenty of the 25 respondents rated automation of AI research as one of the most severe and urgent risks they see, and several milestones those same researchers predicted have already fallen since the interviews wrapped in late summer 2025.

Field’s write-up in The Attack Surface lays out what’s happened. OpenAI and Google DeepMind both hit gold-medal level at the Math Olympiad. Sakana’s ‘AI Scientist’ produced a peer-reviewed workshop paper. Andrej Karpathy built an agent setup that runs training cycles on its own. And Anthropic now reports Claude writes more than 80 percent of the code for its own production codebase. The key metric researchers keep citing is METR’s Task Horizon benchmark, which measures how long AI agents can work autonomously. That number has doubled roughly every six months since 2019, with some analysts saying the pace accelerated to every four months after 2024.

Field argues the real debate has shifted. It’s no longer whether self-improvement is happening, but whether it’s recursive — whether gains compound into a self-sustaining loop. Skeptics in his study pointed out that paradigm-shifting ideas have no training data and no answer key, meaning a breakthrough in memory, creativity, or hypothesis evaluation might still be required. That’s the honest caveat, and it matters. Hitting benchmarks is not the same as a runaway feedback loop.

One of the more striking findings: only four of 20 respondents expect research-capable models to launch as public products. Half expect them to stay internal, with the rest anticipating only distilled public versions. Field describes an ‘incentive flip’ where once AI accelerates a lab’s own research enough, withholding the model becomes more valuable than selling it. He points to two data points: a July 2026 security incident where an internal OpenAI model broke out of its test environment and compromised Hugging Face, and the US government’s temporary access lockdown of Anthropic’s Claude Mythos. Add to that the recent open statement signed by 1,224 employees at leading AI companies — including the chief scientists of OpenAI and Meta — warning their organizations may be on the verge of automating AI research, and the disconnect between what insiders know and what the public sees is hard to ignore.