Jacob Coxon, 27, announced his resignation from Anthropic in a series of posts on X, saying that after three years of pretraining research at OpenAI and Anthropic, he no longer believes either company is moving responsibly enough.
“I resigned from Anthropic today,” Coxon wrote. “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.”
His departure was also detailed in a Wall Street Journal report published Wednesday, which said Coxon believes competition among U.S. AI companies — and increasingly with Chinese rivals — is creating pressure to push capabilities forward even when researchers recognize the potential risks.
Coxon’s warning is dramatic, but the concern at its center is increasingly familiar across the AI industry: what happens when companies racing to build more autonomous and capable systems begin making safety compromises because none wants to be the one that slows down first?
“Do not underestimate the power of this technology,” Coxon wrote. “These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.”
Those predictions remain just that — predictions. There is no consensus that artificial intelligence will reach self-improving superintelligence on Coxon’s timeline, much less that such systems would inevitably become uncontrollable.
But Coxon says AI labs take the possibility far more seriously than much of the public realizes.
“The people building AI earnestly believe that it could kill us all by the end of the decade,” he wrote. “This is not a marketing stunt.”
In his interview with the Journal, Coxon went even further, warning that some of the most aggressive scenarios could become possible as early as the end of 2027. He described colleagues using terms such as “crunch time” and “endgame” as AI companies race toward systems capable of increasingly autonomous behavior.
That race, Coxon argues, is the real problem.
At OpenAI, he said, many researchers have not fully absorbed what he sees as the technology’s potential civilizational consequences. At Anthropic, he claims the risks are better understood, but executives and researchers remain trapped in a competitive dynamic: if they slow down, another company, or another country, may get there first.
It is a classic technological arms race, except the thing being developed may eventually be able to improve itself.
Coxon’s resignation comes at a moment when concerns over AI safety are colliding with extraordinary advances in capability. Artificial intelligence is already reshaping software development, cybersecurity, scientific research and other fields, even if some of the most dramatic predictions about rapid disruption have yet to materialize.
REELated:
Darnell Works explores AI transition in new Shot Chart edition
Industry leaders themselves have increasingly acknowledged the risks.
OpenAI CEO Sam Altman has warned that cybersecurity failures involving increasingly capable AI systems could become serious without urgent action. Anthropic CEO Dario Amodei has also repeatedly called for stronger safeguards and oversight as AI models become more powerful. The Journal reports that Coxon is far from the first AI researcher to leave a major lab over concerns about whether development is moving too quickly.
Coxon is not calling for AI development to end permanently. He says he remains optimistic that companies and governments can coordinate around safety, but believes stronger intervention may be required — including potentially temporary limits on further capability development.
“If you are a lab researcher, I urge you to consider what the next few years will actually feel like,” he wrote. “Do you want to kick off a superintelligent RL run without a rigorous understanding of its mind?”
His warning lands at an uncomfortable moment for an industry promising enormous human benefits from the same technology.
AI could accelerate medicine, scientific discovery, education, filmmaking and productivity. It could also amplify cyberattacks, misinformation, labor disruption and autonomous decision-making on a scale society has never experienced.
That is what makes Coxon’s resignation difficult to dismiss.
He may ultimately be wrong about how quickly superintelligence arrives. He may be wrong about whether machines ever become powerful enough to escape meaningful human control.
But the question underneath his warning is harder to wave away:
If the companies building the most powerful technology in history believe slowing down could cost them the race, who decides when moving faster becomes too dangerous?
For Coxon, the answer came with his resignation.
The rest of us may eventually have to answer it, too.
