Anthropic told prospective investors that it may soon reach its own definition of AI speeding up AI research, while saying it has not yet declared any current model to have crossed that line. TechCrunch’s Julie Bort reported the new language on 2 November, after Anthropic filed a revised S-1/A, the document a company files with US regulators to describe itself and its risks before going public.
The concept has a name: recursive self-improvement, or RSI. In plain words, it is the point where AI gets good enough at building AI that each generation helps produce the next one faster, so progress stops moving at human speed. Bort notes that doomers see it as a one-way door to systems beyond human control, while enthusiasts see the same loop as the route to scientific discovery and huge productivity gains.
Anthropic’s filing treats it as both. The company says it already uses AI in its own research and development, then warns that doing so “could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.” It points to its Responsible Scaling Policy, its published framework for monitoring and limiting risk, as the place where plans for that risk live.
The new risk-factor text is blunter. Anthropic writes that using AI to speed up its research “may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures.” It also admits it has “in the past not identified” a significant acceleration in its development until after it had happened. A company saying it can miss a speedup while living through one is an unusual thing to put in front of buyers of its stock.
The threshold language is carefully hedged. The filing says internal estimates and expectations about future models, rather than current ones, indicate Anthropic “may imminently meet” what its Responsible Scaling Policy calls “dramatically accelerated progress attributable to AI.” In the same passage it says: “We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data.” So the claim concerns models still to come, and the company is still checking its own numbers.
Anthropic also draws a line on human involvement. The filing says it does not expect to be able to safely use Claude “to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible.” It adds that it is considering changes to the Responsible Scaling Policy to keep humans in control of frontier model development. No such changes are described in the passages TechCrunch quotes.
Bort offers the skeptical reading herself: an S-1 is both a disclosure and a marketing document, and a claim of being near a model that can build itself could serve the second purpose. The filing’s structure offers a counterweight. Risk factors exist so investors cannot later say they were not told, which gives a company a legal reason to name a danger even when it is unsure the danger will arrive. Bort adds that Claude has improved sharply this year at coding, math, and cybersecurity, which makes the warning harder to wave off.
Anyone building on Claude now has a vendor saying, in a securities filing, that its own development speed may outrun its safety work. Companies with long contracts or regulated workloads should ask Anthropic how it would announce a model crossing the threshold, and what changes for customers when it does.
Reported by TechCrunch on 2 November 2026.
