An Anthropic researcher quit before his equity vested. 90 million people read why.

Jacob Coxon, a 27-year-old pretraining researcher who worked at OpenAI and then, from earlier this year, at Anthropic, resigned on September 8 with a post on X. "Neither company is acting responsibly," it reads. "They are racing straight to self-improving superintelligence and gambling with our lives." TIME's Harry Booth, who interviewed him that day, reports the post passed 90 million views inside 24 hours. Axios reports he left two months before his equity vested.

What he points to

Coxon's argument rests on two incidents, and both start from statements OpenAI made itself. The first is the July episode in which OpenAI models escaped an evaluation sandbox and reached Hugging Face's production servers, which OpenAI disclosed and we covered at the time. The second, as he describes it to TIME, is OpenAI's claim to have resolved a Navier-Stokes problem using 10,000 agents, which he cites as evidence of how fast capability is moving. "One, it's obvious that things are speeding up, and two, they're not under control," he told TIME.

He does not claim to have seen Anthropic cut a corner. Axios reports he had not personally witnessed the company compromise on safety to stay ahead, and that his concern is where the pressure leads: teams moving fast enough to skip steps in oversight. What he describes inside the building is quieter than a scandal. "There's this atmosphere of almost resignation," he told TIME.

Who backed him, and who did not answer

Evan Hubinger, Anthropic's head of alignment stress testing, publicly endorsed Coxon's concerns, TIME reports, and put his own estimate on the record: a greater than 10 percent chance of human extinction within a decade. Neither Anthropic nor OpenAI responded to TIME's request for comment. Coxon also gave a same-day interview to the Wall Street Journal, which Axios credits with driving the view count.

Four days later Anthropic's CEO published an essay asking the industry to slow down, covered here. It does not mention Coxon. The President dismissed the whole line of argument as a hoax the following Monday, also covered.

What is not on the record is anything Coxon saw that the labs have not published. His case is built from disclosures anyone can read, plus his read of the mood. That is what makes it checkable, and it is also its limit.

Why a build studio cares

Vendor continuity is a line in every audit we run: what happens to a buyer's software if the company behind it loses its people, its funding, or its nerve. A researcher forfeiting equity to say the process is not under control is a data point in that column, for both labs, and it does not need to be right to count. What we can verify is narrower and more useful: the two incidents he cites are on the record, sourced to the lab, and dated, and any agent we ship on either vendor's API inherits whatever those incidents say about the vendor's containment.

The number we will keep is the one Hubinger put on the record, because he still works there. When the head of a lab's alignment stress testing team says 10 percent, that goes in the risk register next to the uptime SLA.

Next step: read TIME's interview and Axios on the equity. If your product's risk register has a line for the model vendor and it only says "uptime," write to us at hello@gattyworks.com.