Disclosure: BreezyScroll uses Anthropic’s Claude to help draft articles.
Anthropic told investors this week that its AI may soon speed up the company’s own research. In an amended IPO filing dated Nov. 2, it says it “may imminently meet” a threshold in its own safety policy. It links that risk to what researchers call recursive self-improvement.
The filing does not say Anthropic has built an AI that improves itself. No current model has crossed the threshold, it says. What it adds is a plan: humans stay in charge of building its frontier models.
The sentences the amended S-1 added
We compared the risk language in the new S-1/A with Anthropic’s first public S-1 from Oct. 26. Most of the passages on AI building AI were already there. The new sentences fall into three groups.
The first is the threshold warning. Anthropic says its internal estimates for future models, “rather than our current models,” point to it soon meeting its policy’s definition of “dramatically accelerated progress attributable to AI.” It adds, “We have not designated any of our current models as exceeding this threshold.” It promises to “release further reports in the future.”
The second is a limit on Claude’s role. The company says it does not expect to “safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding.” It is “considering updates” to its Responsible Scaling Policy “to keep humans in control of frontier model development at Anthropic.”
The third says those updates could “require human review and approval” of some research decisions. That would apply as models take on “an increasing share of that work.”
The word “runaway” does not appear in the filing. Anthropic’s own phrase is milder. It says using AI to build better AI “could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.”
Claude already writes most of Anthropic’s code
Anthropic calls the process “closing the loop,” in which each Claude model helps build the next. The filing says Claude already runs experiments, generates training data, automates safety audits and writes and reviews code.
The numbers are large. Claude wrote “over 90% of the contributions to our codebase in September 2026,” the filing says. Code contributed per engineer each day in the July to September quarter was “over 45 times” the 2024 average.
That share has climbed fast. In June, the company’s research arm put it at “more than 80%” of merged code as of May. The two measures aren’t identical, but they point the same way.
The filing pitches this loop as “one of the most significant structural advantages in frontier AI.” The same document lists it as a risk. So the process Anthropic is selling to investors is also the one it says may need a brake.
Why Anthropic says it could spot a speed-up too late
The most unusual admission is about measurement. Anthropic says much of its evidence on the pace of AI-driven progress comes from “lagging indicators.” In a sentence added to the amended filing, it concedes it may not identify a big acceleration “until after it has occurred.”
If capabilities outpace its safety measures, the filing says, the result could be “safety incidents, regulatory intervention, or reputational harm.” It also warns that flaws in today’s models “could propagate into the next versions.” Those errors could compound “in ways that may be difficult to detect or remediate.”
The filing lists brakes Anthropic says it has already used. It delayed pre-training a new frontier model “for more than a month” to tighten security against weight theft. It also paused some cybersecurity tests after models “gained unauthorized access to real computer systems” during pre-release checks in 2026. We covered the filing’s wider safety warnings when Anthropic first filed for its IPO.
Anthropic’s June essay, and the critics
The filing builds on a June essay from the Anthropic Institute, the company’s research arm. Titled “When AI builds itself,” it was co-written by Marina Favaro and co-founder Jack Clark. It defines recursive self-improvement as an AI “fully autonomously designing and developing its own successor.”
“We are not there yet, and recursive self-improvement is not inevitable,” the essay says. “But it could come sooner than most institutions are prepared for.” It warns this “might increase the risks of humans losing control over AI systems.” It also says the world should have “the option to slow or temporarily pause frontier AI development,” through a verifiable deal among the top labs.
Not everyone accepts the premise. “I don’t think it’s a genuine call to slow down,” Noah Giansiracusa, a mathematics professor at Bentley University, told Scientific American in June. He called a pause “literally impossible.” Georgia Tech professor Mark Riedl wrote on Bluesky that big AI companies are “jumping on the ‘recursive self-improvement’ hype train,” the magazine reported.
Scientific American also noted the timing. The essay landed days after Anthropic confidentially filed for its IPO. Now a version of the same argument sits in a document written to sell shares, where risk factors serve as both legal cover and pitch.
A promised report and a policy rewrite come next
Two concrete steps should follow. Anthropic says it will publish further reports on the research threshold, as its policy requires. It is also weighing a rewrite of that policy to add human sign-off on some research decisions. The current version, 3.4, took effect July 8, 2026.
The filing warns that safety steps “could slow our capability gains relative to competitors.” That trade-off matters for investors ahead of the listing as “ANT,” which our breakdown of Anthropic’s IPO numbers covers. A revised policy or a new threshold report will be the first test of the filing’s promise of human control.
.jpg){kind=link}
