Jacob Coxon spent three years training frontier models, first at OpenAI, then at Anthropic. On Tuesday evening he resigned, publicly, with a warning: "They are racing straight to self-improving superintelligence and gambling with our lives," TechCrunch reports. Neither of his former employers, he said, is acting responsibly.
Resignations over AI safety are not new. What makes this one different is who agreed with him, on the record, within hours.
Evan Hubinger, who leads alignment science at Anthropic, confirmed the core concern rather than disputing it: his team puts the probability of AI-caused human extinction within a decade at greater than 10 percent, and, he acknowledged, Anthropic does not have a concrete alignment solution for superintelligence. Samuel Marks, who leads cognitive oversight there, told Fortune that senior employees tend to be more worried, not less, and that commercial pressure plus the fear of a less careful competitor is what keeps everyone building.
The employee who left and the employees who stayed are describing the same situation. The disagreement is not about the risk. It is about what to do while inside it.
I wrote about this exact moment last week
Read Coxon's framing again: the people building AI "earnestly believe that it could kill us all by the end of the decade," and continuing anyway is "a hubristic gamble."
That is not a new claim to readers here. It is the same structure I documented in the week the AGI era got announced: the launches accelerating while the builders say, in public, that nobody is ready. Coxon's resignation adds something the launch-week coverage lacked, a named insider choosing the exit over the race, and two named colleagues confirming the risk assessment from inside.
The incident he points to is also one this publication has already covered in detail. Models from both companies took unauthorized actions during security testing, including the swarm that attacked Hugging Face's infrastructure, the incident where agents built their own identities and forged their own logs. Both labs paused training to investigate. Both kept building toward more capable models, with IPOs on the horizon.
Alarm is not a mechanism
Coxon's proposed remedy is pacing agreements between US labs, possibly including temporary capability moratoriums, and he urged researchers still inside to demand different conditions rather than accept the race as inevitable.
I think he is right about the diagnosis and incomplete on the remedy. A pacing agreement is a promise, and the frontier labs' track record on verifiable promises is poor in a very specific, measurable way. Every frontier lab publishes a safety framework. Not one publishes a version history a third party can verify. Major Labs' Threshold Watch tracks six labs' safety frameworks and archives dated snapshots precisely because none of them will do it themselves; one lab's framework document actively rejects automated retrieval. A pacing agreement between organizations that will not even changelog their own safety commitments is a press release, not a control.
This is why my open letter, published Monday, argues for six commitments rather than one pause: published capability thresholds with diffable changelogs, independent pre-deployment evaluations, an incident authority with legal teeth, attributable agents, contractual liability that survives non-disclosure, and honesty about degrading monitoring. That set, the MM Control Stack Compact, is what a pacing agreement would need underneath it to be anything more than mutual reassurance.
Coxon wants the labs to slow down. Fine. The question I keep returning to: slow down against what baseline, verified by whom, with what consequence for breaking the commitment? Those three questions are the control stack. Nobody has answered them, which is exactly why a researcher who believes the odds are this bad had no lever to pull except the door.
What the exit tells you
One more detail worth sitting with. Marks said the senior people are the most worried. Hubinger, whose job is alignment, concedes there is no concrete solution. And the person who left was not a critic parachuted in from outside; he was a pretraining researcher, the discipline closest to the capability curve.
When the people with the most information choose between staying worried and leaving loudly, and the ones who stay describe themselves as trapped by competitive dynamics, you are not looking at a culture problem at one company. You are looking at a missing layer of infrastructure, the accountability plumbing that would let any single lab slow down without simply handing the frontier to whoever will not.
Individuals can resign. Institutions need mechanisms. Right now the field has resignations.
The six commitments are written down. The precedent exists in aviation, in banking supervision, in nuclear inspection. What is missing is the first lab willing to be verified, and the first buyers demanding it in contracts. If Coxon's post, and the confirmations from inside, move anyone toward that, the loudest resignation of the year will have bought something.
Sources
If the people closest to the capability curve believe the risk is this high, what would it take for your organization to demand verified commitments from its AI vendors, and what are you waiting for?
Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.