Anthropic tells the SEC its AI may soon be speeding up its own development

IPO filings are where companies list everything that could go wrong, written by lawyers whose job is to make sure no shareholder can later say they weren't warned. Most risk factors are boilerplate about interest rates and key employees. Anthropic's amended S-1, filed with the Securities and Exchange Commission on Monday, has a new one that reads like the opening of a science fiction novel.

According to TechCrunch, which first reported the change, Anthropic now tells prospective investors that "we may imminently meet our Responsible Scaling Policy's definition of" a capability threshold it describes as "dramatically accelerated progress attributable to AI." In the industry this idea goes by the name recursive self-improvement, or RSI: AI systems that get good enough at AI research that each generation speeds up the building of the next.

What the filing actually says

The wording is hedged, as you'd expect. Anthropic says it has not designated any current model as crossing that line, and that it is still assessing the data. The warning is about models it expects to build, based on its internal estimates. Dataconomy, which also reviewed the filing, reports the same.

The company is open about the fact that it already uses AI in its own research and development. It also says, per TechCrunch, that "using AI to help build more capable AI could also pose a considerable safety risk." Dataconomy quotes the filing's concern that this could cause "model capabilities to advance faster than its ability to develop, validate and implement" the safety, alignment and security measures meant to keep up with them.

The most uncomfortable line is about Anthropic's own track record. The filing says "we have in the past not identified, and may in the future not identify, a significant acceleration" in its rate of progress until after it had happened. TechCrunch reads this as an admission that Anthropic has already used AI to build models whose jumps in capability it did not recognize at the time.

That's a striking thing for a company to put in a legal document. The whole premise of a scaling policy like Anthropic's is that you measure capabilities as you go and add safeguards before you cross a dangerous line. The filing says the measuring has, at least once, run behind the thing being measured.

Keeping a human on the loop

Anthropic's answer is a promise about supervision. "We do not anticipate being able to safely use Claude to fully autonomously design and build successive generations" of frontier models, the filing says, according to TechCrunch. Dataconomy reports the rest of that thought: Anthropic doesn't expect to do this safely without meaningful human supervision or understanding, even if it becomes technically possible. The company says it wants to "keep humans in control of frontier model development at Anthropic," and that "accordingly, we are considering updates to our Responsible Scaling Policy" to back that up.

Note the tense. Considering updates. The policy that is supposed to govern this moment is, by the company's own account, still being rewritten as the moment arrives.

There is a fair cynical reading of all this, and TechCrunch's Julie Bort raises it: an S-1 is both a disclosure and a sales document. Telling investors your technology is so powerful it might start building itself is a warning, and it's also a pitch. Anthropic has spent years arguing that frontier AI is dangerous enough to need careful labs like Anthropic. A risk factor that says "our models may soon accelerate their own development" lands on both sides of that argument at once.

The legal logic points the other way, though. Securities lawyers put things in risk factors because leaving them out is the bigger liability. If Anthropic believed RSI was distant and speculative, the safest move would be to say nothing specific. Instead the filing uses the word "imminently" and admits to past blind spots. Someone at the company decided the exposure from staying quiet was worse.

For readers outside the AI world, the practical question is what "dramatically accelerated progress" would look like from the outside. Probably not much at first. Faster release cycles. Bigger jumps between versions. Research papers where the AI did more of the work. TechCrunch notes Claude's gains in coding, math and cybersecurity this year as context, and those are exactly the skills an AI would need to work on itself.

The filing doesn't say when. It says soon, it says the company might not notice right away, and it says the rules are under review. Those are admissions in a document Anthropic's lawyers combed line by line before it went to the SEC. Investors were warned. So, in a roundabout way, was everyone else.

Sources: TechCrunch, Dataconomy.