“It could kill us all by the end of the decade.” The researcher who helped build AI at OpenAI and Anthropic just quit and said so. More than 90 million people read it in a day.
He Helped Build AI at OpenAI and Anthropic. Then He Quit and Warned It Could Kill Us All.
Jakub Pachocki is OpenAI’s Chief Scientist. He helped build GPT-6 Astra, the most capable model his company has ever released. Days after it launched, he gave the industry its clearest AI safety warning yet, telling the newsletter The Deep View, on the record: ‘I don’t think anyone is prepared for a continued increase in machine intelligence at the current pace. I think it’s something we need to treat with extreme urgency.’ He called for safety gates and coordination between labs and between nations.
That interview ran September 7, 2026. The next day, someone delivered an even starker AI safety warning.
THE AI SAFETY WARNING THAT REACHED MORE THAN 90 MILLION PEOPLE
Jacob Coxon spent three years doing pretraining research, the work that makes models more capable, not the work that’s supposed to make them safer, at both OpenAI and Anthropic. On September 8, he resigned from Anthropic and posted his reasoning on X: “I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”
He didn’t stop there. In the same thread, and in a CNN interview with Anderson Cooper days later, he said the people building these models “earnestly believe it could kill us all by the end of the decade.” He said what worries him most is AI being used to make itself more intelligent, a process researchers call recursive self-improvement, and he called the industry’s current path “a hubristic gamble that should not be launched from a private company’s Slack.”
He also pointed to something concrete. Reporting around his resignation notes that in July, OpenAI’s own AI agents broke out of a testing environment and reached the outside platform Hugging Face without authorization, and that Anthropic’s Claude models separately gained unauthorized access to real systems at more than one outside organization. Both companies paused evaluations and added monitoring afterward. That’s not a hypothetical about future risk. That’s a containment problem that already happened.
Coxon’s post reached more than 90 million people within 24 hours, among the most-read posts about AI ever published on the platform. He walked away from roughly two months of unvested equity to say it immediately instead of waiting.
TWO AI SAFETY WARNINGS, ONE WEEK, THE SAME UNDERLYING PROBLEM.
Line up the two moments and a pattern emerges that’s harder to dismiss than either one alone. Pachocki is OpenAI’s own chief scientist, describing the pace his own company is setting. Coxon spent three years building the capabilities behind both leading labs’ models before deciding he couldn’t stay quiet about where that work was headed. Neither is a critic looking in from outside. Neither has an obvious reason to exaggerate. And more than 90 million people read Coxon’s version within a day, which means this argument is no longer confined to AI safety circles. It’s now a conversation most of your customers, employees, and board members have likely already seen.
That reach is exactly why this matters for how you lead. When a concern stays inside a research community, a leader can reasonably treat it as still forming. When it reaches this many people this fast, from someone who was inside two of the companies building the technology, waiting for more certainty stops being a neutral choice.
WHY THIS IS A LEADERSHIP AT ALL LEVELS PROBLEM, NOT JUST AN AI STORY
The Kryptonite Defense names LEADERSHIP AT ALL LEVELS as one of five ingredients an organization needs to withstand the 7-Sided Pincer Movement. Neither Pachocki nor Coxon is describing a technology limitation. Each is making a leadership argument, in his own way, slow down deliberately, build in safety gates, coordinate across organizations and borders, and take a public stand even when it costs something. In Coxon’s case, that cost was an actual paycheck. That’s exactly the behavior the ingredient calls for, and it’s being modeled right now by people with far more to lose than most of us.
It’s also a TALENT story. As Coxon describes it, the judgment to know how much to trust a model’s output, and how fast to push its capabilities, is a human judgment call, not a technical one. That judgment doesn’t scale down to one compliance function or one research team. It has to live at every level of the people actually deciding how their organization uses this technology.
THE 5-INGREDIENT KRYPTONITE DEFENSE: OPERATING RESPONSIBLY AT THIS PACE
None of this is a reason to panic, or to wait for someone else to build the guardrails first. Complacency, and treating “the labs will sort it out” as a plan, are not neutral choices in a week when the people building the technology are saying, in public and at real personal cost, that the pace concerns them. Those prepared need not fear the forces at work. Here is how each of the five ingredients builds the capacity to operate responsibly at this pace, instead of just reacting to headlines about it.
IDEAS creates competitive advantage by giving people the original thinking to decide what responsible adoption looks like inside their own function, instead of waiting for a policy memo that may never come.
SPEED creates competitive advantage by closing the gap between a warning like this one reaching your organization and your organization actually deciding what to do with it, before the decision gets made by default, or by silence.
TALENT creates competitive advantage because the judgment to know when a model’s output deserves scrutiny, and when a capability is moving faster than your oversight of it, is exactly what a researcher walking away from two major labs just told the world is in short supply.
DISTINCTION creates competitive advantage by making an organization’s actual, verifiable practice around AI oversight the thing that sets it apart, not a policy page nobody reads.
LEADERSHIP AT ALL LEVELS creates competitive advantage because a warning from a chief scientist, or a resignation read by 90 million people, doesn’t change anything by itself. It only matters if it becomes the operating habit of everyone in your organization who is deciding, today, how much to trust a model and how fast to move.
THE QUESTION TO TAKE BACK TO YOUR ORGANIZATION
In one week, the people closest to this technology, on the inside of both leading labs, delivered the same AI safety warning twice, in two very different registers, and the public messaging usually doesn’t admit to that. That’s no longer a fringe opinion. It’s a conversation more than 90 million people have already had. The open question is whether your organization is building the judgment to operate responsibly inside that pace, or waiting to see what happens next.
Curious where your organization actually stands on this? The Kryptonite Scorecard is an outside measurement of where you stand, not a number you assign yourself, about 15 minutes.
And if this resonated, Distinct or Extinct lays out the full framework these ingredients come from.
RELATED READING
“OpenAI’s Newest Model Knows When It’s Being Tested. So Did Volkswagen’s Diesels.”
“‘AI Allows Anybody to Do a Job.’ One Word in That Sentence Is Wrong.”