After CEO Dario Amodei's repeated warning that AI will wipe away millions of jobs, Anthropic publishes a 10,000-plus word paper to tell everyone AI can be more dangerous than just taking jobs, it can also…

Anthropic spent most of 2026 warning the world that AI was coming for white-collar jobs. CEO Dario Amodei said it at Davos, in essays, at the India AI Impact Summit, and on most podcasts that would have him. Unemployment could hit 10 to 20% inside five years. Coding goes first, then finance, then law. Pension funds and sovereign wealth managers, weighing a reported $900 billion valuation for Anthropic, heard a market the size of the global wage bill. Now the company has a new message, and it sits awkwardly on top of the old one.

In a paper that runs past 10,000 words, titled When AI builds itself, Anthropic argues that the bigger danger isn't a wave of laid-off coders and junior analysts. It's the prospect of an AI system capable of designing and training its own successor, with humans pushed to the edges of a process they used to run from start to finish. The company calls this recursive self-improvement. It says we aren't there yet. Then it spends thousands of words laying out evidence that the gap is closing fast, and warns that if it does close, the work of securing, monitoring, and shaping these systems becomes far more urgent than anyone is currently treating it.

Inside Anthropic, Claude is already writing the company that builds Claude

The paper, co-authored by Marina Favaro and Jack Clark of the Anthropic Institute, leans on internal data the company has never disclosed before. As of May 2026, more than 80% of the code merged into Anthropic's own codebase was authored by Claude. Before Claude Code launched in February 2025, that share sat in the low single digits. The average Anthropic engineer now ships 8x as much code per quarter as they did between 2021 and 2025. In a March 2026 internal poll, the median researcher said they were producing roughly four times as much output with the still-internal Mythos Preview model as they would have without any AI at all.

One example the paper offers is hard to wave away. In April 2026, Claude shipped over 800 fixes that cut a class of API errors by a factor of one thousand. The engineer overseeing the work estimated a human would have taken four years to finish it. External evaluator METR reports that Mythos Preview can now work on its own for at least 16 hours, sitting at the upper end of what the lab can even measure. The length of tasks AI can complete reliably is doubling every four months, up from seven last year.

The risk Anthropic is now flagging isn't unemployment, it's losing the steering wheel

This is where the framing pivots away from Amodei's job-loss circuit. Anthropic argues that if an AI system becomes capable of building its successor with no meaningful human input, alignment stops being a research problem and becomes a survival one. The paper warns that the rare cases of misalignment present in today's models could compound as those models build the next generation, growing more frequent but less understood until humans lose control of them.

The company also hints at how fast things move once an AI is loose at scale. Project Glasswing, Anthropic's program that gives Mythos Preview to a small group of trusted partners, surfaced more than 10,000 high- and critical-severity software vulnerabilities across major systems in its first few weeks. The bottleneck has already shifted from finding holes to patching them fast enough.

Why Anthropic is now floating the idea of a coordinated pause on frontier AI

For a company chasing a reported $900 billion valuation, the paper's closing section is the most striking part. Anthropic says it would support a slowdown or temporary pause on frontier AI development, on one condition: that other top labs do the same, in a way that can be independently verified. The company concedes this is hard. Training runs are easier to hide than missile silos. The incentive to defect is enormous. There is no global body to adjudicate. But Anthropic says the Anthropic Institute will start building the verification systems a credible pause would need, and will convene policymakers, researchers, and other AI companies in the coming months.

For now, Claude still waits for Anthropic engineers to tell it what problem to solve. The paper's argument is that the gap is closing, and that the public conversation has spent too much time on whose job goes first, and not enough on who, or what, is in charge when the process starts running itself.