An Anthropic Researcher Just Quit Warning AI Could "Kill Us All" by the End of the Decade — Here's What He Actually Said

Jacob Coxon, a 27-year-old researcher who spent three years on pretraining work at OpenAI and then Anthropic, resigned this week with a seven-part thread on X that drew roughly 76 million views overnight, plus an exclusive interview with the Wall Street Journal.

His claim, in his own words: the labs are "racing straight to self-improving superintelligence and gambling with our lives." He told the Journal that "we're on track for a lot of the most aggressive of these scenarios where by the end of next year things could be out of control already," and that researchers inside frontier labs increasingly use the words "crunchtime" and "endgame" to describe where capability progress is heading.

Coxon moved from OpenAI to Anthropic earlier this year specifically because Anthropic has a reputation for taking safety more seriously. He left disappointed, writing that "neither company is acting responsibly." He also said the thing people actually want is regulation — the New York Post reported him saying the public is "begging" for it.

Why this matters for people who use these tools

This is not a prediction that the world ends. It is a data point, and a loud one: a senior insider is saying, publicly and on the record, that the timeline the companies market and the timeline researchers talk about among themselves are not the same timeline.

Three concrete things follow from that for anyone whose work now depends on these tools:

1. Regulation is coming, and it will touch features. When an ex-safety researcher goes on the record and 76 million people see it, that is the kind of moment that moves legislators. The last two years of "ship first, ask later" is the part most likely to get constrained — autonomous agents that take actions, models that can modify their own code, anything that runs without a human in the loop. If your workflow leans on those, expect friction.

2. The "safety versus speed" fight will decide which tools last. Coxon's whole point is that the two big labs disagree internally about how fast is too fast. That disagreement plays out as product decisions: what gets released, what gets held back, what gets a capability cap. The tool you standardise on today may not behave the same in a year.

3. "Just add an agent" deserves a harder look. The advice all over social media right now is to wire an autonomous agent into everything — your shell, your inbox, your deploys. The people building the underlying models are using the word "endgame." That is not a reason to stop using AI tools. It is a reason to keep a human approval step on anything an agent can do that you cannot easily undo.

The honest caveat

Coxon is one researcher, he is 27, and resignation threads are written to be shared. Evan Hubinger and others named alongside him have not all echoed the same framing. Treat the timeline as one insider's read, not a consensus. But "one insider's read" from someone who did pretraining at both leading labs is still worth more than a marketing blog post, and it points the same direction as the regulation debate already underway.

The tools are not going away. The rules around them are about to change, and the people closest to the technology are the ones asking for it.

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →