Jacob Coxon, who worked at OpenAI and Anthropic, said both companies are moving toward self-improving superintelligence too quickly. He said researchers within the industry have serious concerns that such AI could pose a threat to humanity.
Jacob Coxon, who worked at OpenAI and Anthropic, said both companies are moving toward self-improving superintelligence too quickly. He said researchers within the industry have serious concerns that such AI could pose a threat to humanity.
Researcher Jacob Coxon has announced his departure from Anthropic and said he is leaving the AI industry. He said he has spent the past three years doing pretraining research at OpenAI and Anthropic.
Coxon said that both companies, in his view, are moving towards creating self-improving superintelligence and are doing so without adequate safeguards. He called the situation a «race» and said that the companies are «playing with our lives.»
In a series of posts after his release, Coxon wrote that the people who create modern AI systems are seriously considering a scenario in which artificial intelligence could destroy humanity by the end of this decade.
He stressed that this, he said, is not a public marketing message. Coxon claims that some executives and senior researchers express much more serious concerns in private conversations than they do publicly.
Coxon also urged not to underestimate the capabilities of future AI systems. He expects them to be able to perform tasks that currently require human expertise, access resources, and work across domains much faster than humans.
According to him, some researchers are already using the words «crunchtime» and «endgame» to describe the current stage of AI development. In an interview with The Wall Street Journal, Coxon said that under the most aggressive scenarios, the situation could get out of control by the end of 2027.
Coxon worked at OpenAI from 2023 to July 2026. According to Business Insider, he worked on GPT-4o, among other things. In July, he moved to Anthropic as a researcher.
Coxon says he joined Anthropic because the company is known for its focus on AI security. However, he later came to the conclusion that even Anthropic cannot solve the problem of superintelligence security on its own while competition between AI labs continues.
Coxon describes the difference between the two companies as follows: he believes that at OpenAI, some employees are not fully aware of the potential consequences, while at Anthropic, these risks are well understood. At the same time, he says, Anthropic continues to move forward because it believes that if it doesn’t build powerful AI first and try to do it responsibly, someone else will.
Coxon believes that solving this problem requires coordination between AI labs, and perhaps a temporary ban on further expanding the capabilities of models.
He also urged other researchers to consider whether they are ready to launch training on superintelligent systems without having a sufficient understanding of how these systems work and how to control their behavior.
After Coxon’s publications, Evan Gubinger, who heads the Alignment Science department at Anthropic, responded to them.
Gubinger said he agreed with Coxon and that he and his colleagues do believe that a scenario in which AI could wipe out humanity is possible. Gubinger personally estimates the probability of such a scenario to be more than 10% within the next decade.
At the same time, he separately clarified that he considers the risk from current models to be low. His concerns relate to another scenario — the emergence of superintelligence as a result of recursive self-improvement, that is, the repeated self-improvement of the system. According to Gubinger, this process may occur faster than previously expected.
Gubinger also stated that Anthropic is trying to solve the problem, but the company does not yet have a plan that would solve the alignment problem for superintelligence, and is not currently on a clear path to such a solution.
Coxon and Gubinger’s statements came amid a debate about whether the development of the most powerful AI systems should be slowed down.
In July 2026, 1,386 employees of frontier AI companies signed the Pacing the Frontier initiative. The signatories included employees of OpenAI, Anthropic, Google, Meta, and others. They called on the US government to support international mechanisms that would allow for a controlled slowdown in the development of frontier AI, if necessary.
One of the reasons for the discussion was the capabilities of AI agents that can perform complex sequences of actions without constant human control. The participants of the initiative separately spoke about the future automation of the AI research process itself — a situation where AI will begin to help create the next generations of AI.
Coxon believes that the transition to systems capable of independently improving their own capabilities is a fundamentally new stage of risk.
At the same time, his statements and Gubinger’s assessment are their predictions and risk assessments, not confirmation that modern AI systems are already capable of destroying humanity or that such a scenario will definitely happen.
