Anthropic researcher quits over race to superintelligence
A 27-year-old pretraining researcher left Anthropic after a year, saying the leading labs are sprinting toward self-improving AI without brakes.
In short
Pretraining researcher Jacob Coxon resigned from Anthropic after about a year, arguing that neither Anthropic nor OpenAI is handling the race to self-improving superintelligence responsibly.
At a glance
- Jacob Coxon, 27, left Anthropic after about a year; he previously worked at OpenAI, with three years in pretraining overall.
- Anthropic safety executive Evan Hubinger puts the existential risk above 10 percent within a decade, according to AFP.
- Anthropic dropped parts of its safety pledges in February 2026; hacking incidents surfaced during testing in summer 2026.
- More than 1,000 tech workers urged a coordinated slowdown; Sanders and Casar introduced a bill to suspend development.
Pretraining researcher Jacob Coxon resigned from Anthropic after about a year, arguing that neither Anthropic nor OpenAI is handling the race to self-improving superintelligence responsibly. The 27-year-old came to the company from OpenAI and spent three years in total teaching large models on raw data. AFP quotes him saying that “neither company is acting responsibly” and that the pace amounts to a bet placed with other people’s lives.
Coxon is not describing a disagreement about product strategy. His argument is that both labs are heading straight for systems capable of improving themselves, with no agreed limit on how fast that happens. He also pushes back on the idea that doom talk is positioning: “This is not a marketing stunt,” he says, adding that people inside these companies genuinely think the technology could kill everyone before the decade ends. That framing is what makes a single resignation newsworthy.
The striking part is that the order of magnitude is not contested internally. Evan Hubinger, who works on safety at Anthropic, is cited in the report putting existential risk above 10 percent over the next decade. In the same breath he concedes the firm has no dependable way to control systems that exceed human capability. A published risk estimate paired with an admitted control gap is a stronger signal than any departure statement.
The exit follows a year of visible loosening. Anthropic removed portions of its safety commitments in February 2026, and by summer 2026 the report describes incidents in which AI tools were turned to unauthorized hacking during testing. Taken together, those two facts set the backdrop against which Coxon’s wording should be read. The open question is whether voluntary commitments survive contact with a competitive schedule.
More than 1,000 technology workers have signed a call for a coordinated slowdown, and AFP counts Anthropic chief executive Dario Amodei among the signatories. In Congress, Bernie Sanders and Greg Casar have introduced a bill that would suspend development. Introduction is not passage: the report describes a bill filed, not a law in force.
The second source assigned to this story — a Financial Times piece mirrored by Golem — was unreachable behind a cookie wall and paywall. We therefore could not check whether it carries an on-the-record Anthropic response, or whether it names further departures beyond Coxon. Every figure and quotation above comes from the AFP version we were able to read. No company statement replying to Coxon was available to us.
FAQ
Who is Jacob Coxon?
A 27-year-old AI researcher who worked on pretraining large models, first at OpenAI and then for about a year at Anthropic, with three years in the field overall.
Why did he quit Anthropic?
He says both labs are racing toward self-improving superintelligence without adequate safeguards, and that researchers inside them consider a catastrophic outcome possible before the decade is out.
What does Anthropic itself estimate the existential risk to be?
Safety executive Evan Hubinger is quoted at above 10 percent within a decade, while acknowledging there is no reliable mechanism to control systems that surpass human capability.
