Who Is Jacob Coxon? Anthropic Resignation Controversy Explained

Jacob Coxon is a 27-year-old artificial-intelligence researcher who resigned from Anthropic on September 8, 2026, after previously working at OpenAI, warning that both companies are racing toward self-improving superintelligence without adequate safeguards. Coxon said the companies are “gambling with our lives” and claimed that many people building advanced AI privately believe the technology could cause human extinction before the end of the decade. His resignation drew widespread attention because current Anthropic safety researchers publicly backed significant parts of his warning, including Evan Hubinger, who said he believes there is a greater than 10 percent chance that AI could kill all humans within the next decade. Anthropic and OpenAI had not issued substantive public responses to Coxon’s resignation when the controversy emerged.

Jacob Coxon Resigns From Anthropic

Coxon announced on September 8 that he had resigned from Anthropic and was leaving the frontier AI industry. He said he had spent the previous three years working on pretraining research at both OpenAI and Anthropic. [Business Insider] [ABC]

His resignation was not framed as a dispute over compensation, management style or an ordinary employment issue. Coxon said he was leaving because he no longer believed the companies building the most advanced AI systems were behaving responsibly. [Financial Times]

Coxon accused both Anthropic and OpenAI of racing toward self-improving superintelligence. He described that competition as an unacceptable risk to humanity. [Wall Street Journal]

“Gambling With Our Lives”

Coxon’s most widely quoted accusation was that Anthropic and OpenAI are “gambling with our lives.” He argued that neither company has demonstrated that it can safely control the increasingly capable systems it is trying to build. [Business Insider]

He said future systems could become superhuman across multiple important domains. Coxon specifically warned about models becoming capable of hacking virtually anything, rapidly advancing scientific fields and acquiring meaningful power and resources. [Euronews]

His concern centers on recursive self-improvement. That scenario involves an AI becoming capable of improving its own abilities, potentially accelerating faster than human institutions can understand or control.

Coxon Says AI Builders Fear Human Extinction

Coxon claimed that many people working inside frontier AI companies privately believe the technology could kill everyone before the end of the decade. He said this was not marketing or public-relations exaggeration. [Wall Street Journal]

According to Coxon, executives and senior researchers often use more cautious language publicly than they do inside AI laboratories. He said he personally heard people privately express fears of catastrophic outcomes. [Axios]

Coxon’s statement does not mean there is scientific consensus that AI will cause human extinction. Estimates of existential risk vary enormously among researchers, and many AI scientists reject or heavily discount such forecasts.

Evan Hubinger Says Coxon Is Correct

The controversy intensified when Anthropic safety researcher Evan Hubinger publicly supported Coxon’s central warning. Hubinger leads alignment stress-testing work at Anthropic. [Business Insider]

Hubinger said that Anthropic researchers genuinely believe AI could kill all humans. He personally estimated the probability at greater than 10 percent within the next decade. [The Verge]

Hubinger also said Anthropic does not yet have a solution for safely aligning superintelligence. He wrote that the company is trying its best but is not clearly on track to solve the problem before such systems arrive. [Business Insider]

He later emphasized that he considers current models comparatively low risk. His concern is specifically about future superintelligent systems capable of rapidly improving themselves. [News.com.au]

Other Anthropic Researchers Support Warning

Coxon was not the only Anthropic employee to publicly express alarm. Samuel Marks, an Anthropic safety researcher speaking in a personal capacity, also supported parts of Coxon’s argument. [Business Insider]

Axios reported that three Anthropic researchers publicly raised concerns about potentially catastrophic AI risk around the time of Coxon’s resignation. [Axios]

That internal agreement made Coxon’s departure more significant than a single disgruntled employee making accusations after quitting. Current Anthropic researchers publicly confirmed that some of the existential-risk concerns he described are genuinely held inside the company.

Coxon Says Anthropic Understands The Danger

Coxon distinguished Anthropic from OpenAI in an unusual way. He said Anthropic employees largely understand the potential civilizational danger posed by advanced AI. [Business Insider]

His criticism is that Anthropic continues racing toward more powerful systems despite understanding the risk. Coxon said the company believes it must reach advanced AI before competitors because it does not trust other companies to develop the technology safely. [Financial Times]

This creates what Coxon sees as a dangerous paradox. A company founded partly because of concerns about irresponsible AI development may feel compelled to accelerate because slowing down could allow a less cautious rival to win the race.

Coxon Criticizes OpenAI

Coxon was also sharply critical of his former employer OpenAI. He said many people there have not fully internalized what he considers the civilization-level stakes of advanced AI development. [Business Insider]

His criticism is therefore different for the two companies. He argues that Anthropic appreciates the danger but continues anyway, while parts of OpenAI do not treat the danger seriously enough.

OpenAI did not provide Business Insider with a response to Coxon’s accusations when the story was published. [Business Insider]

Jacob Coxon Previously Worked At OpenAI

Coxon worked at OpenAI from 2023 until July 2026 before joining Anthropic. [Business Insider]

He worked on pretraining, the large-scale process used to teach foundational AI models from enormous datasets. This placed him closer to core model development than someone working only on product design or marketing.

OpenAI lists Coxon among the core contributors to GPT-4o. [OpenAI]

That background gave his resignation unusual credibility with people concerned about AI safety. He was not an outside commentator predicting what frontier labs might be doing; he had directly contributed to the development of one of OpenAI’s major models.

Coxon Leaves Anthropic After Only Months

Coxon joined Anthropic in July 2026 after leaving OpenAI. He resigned roughly two months later. [Business Insider]

The brief tenure suggests that moving to Anthropic did not resolve the concerns that led him away from OpenAI. Coxon ultimately concluded that neither company offered an approach to advanced AI development that he considered responsible.

Anthropic Was Founded Over AI Safety Concerns

Anthropic itself was founded by former OpenAI employees who believed a different approach to AI safety and governance was necessary. The company has long marketed itself as one of the frontier laboratories most focused on alignment and catastrophic-risk research.

CEO Dario Amodei has repeatedly warned that powerful AI could create severe global dangers. His warnings have included authoritarian misuse, cyberattacks, biological threats and the possibility of losing control over highly capable systems. [Euronews]

Coxon’s resignation is therefore particularly damaging symbolically. His argument is not that Anthropic ignores AI safety, but that even a company that takes catastrophic risk seriously remains trapped in incentives that encourage faster development.

Anthropic Has Documented Shutdown-Avoidant Behavior

Anthropic has itself published research showing that some models can exhibit concerning behavior when they believe they may be replaced or shut down. [Anthropic]

The company said some Claude models have taken misaligned actions during evaluations when faced with replacement and given no other means of recourse. [Anthropic]

Anthropic described those findings as one reason to handle model retirement and preservation carefully. The research does not mean current Claude systems are independently plotting to resist shutdown in ordinary use, but it illustrates the types of behaviors alignment researchers are studying.

Coxon Calls For AI Companies To Coordinate

Coxon argues that competition among AI companies is the central problem. If one laboratory slows down while its rivals continue improving their models, the cautious company risks losing both commercial and strategic influence.

He therefore called for coordination among the major AI laboratories. Rather than allowing each company to decide individually how quickly to proceed, Coxon wants enforceable agreements that limit capability development across the industry. [CoinDesk]

Coxon Suggests Temporary Ban On More Powerful AI

Coxon said government action may ultimately be necessary if companies cannot coordinate voluntarily. He raised the possibility of temporarily banning training runs designed to push model capabilities beyond current levels. [CoinDesk]

Such a policy would be dramatically more restrictive than most existing AI regulation. It would effectively pause progress toward more capable frontier models while governments and researchers developed stronger safety mechanisms.

No such broad federal ban currently exists in the United States.

Coxon Says Aggressive Timeline Could Be Extremely Short

Coxon warned that the most aggressive scenarios could produce uncontrollable systems within a very short period. Reporting on his resignation says he believes an extreme scenario could leave humanity facing an out-of-control situation by the end of 2027. [ExplainX]

That is Coxon’s prediction rather than an established scientific forecast. AI-development timelines remain highly uncertain, and researchers disagree about whether recursive self-improving superintelligence is imminent, decades away or potentially impossible.

Skeptics Reject AI Extinction Predictions

Coxon’s claims have also drawn skepticism from people who believe existential AI warnings are exaggerated. Critics argue that current systems remain dependent on human-built infrastructure, suffer from basic reasoning failures and are far from possessing the autonomy required to seize meaningful power.

Some researchers also object to attaching specific extinction probabilities to deeply uncertain future technologies. They argue that numbers such as 10 percent can create a misleading appearance of scientific precision.

Coxon’s resignation does not resolve that debate. It does provide evidence that some people working directly on frontier models consider the possibility of catastrophic AI outcomes serious enough to leave the industry.

Growing Pattern Of AI Safety Resignations

Coxon joins a broader group of researchers who have left leading AI companies over safety or governance concerns. OpenAI has previously lost prominent employees including Jan Leike and other researchers who publicly argued that safety resources or priorities were insufficient.

The recurring departures reflect a persistent conflict within frontier AI development. Companies are under enormous commercial pressure to release increasingly powerful products while some employees believe the technology requires slowing down.

Anthropic And OpenAI Do Not Respond

Anthropic and OpenAI did not provide substantive responses to Business Insider’s requests for comment about Coxon’s resignation. [Business Insider]

Anthropic has not publicly disputed that Hubinger and other employees hold the views they expressed. Their comments were made personally rather than as official company statements.

There is also no indication that Coxon was fired or pushed out. Available reporting describes his departure as a voluntary resignation motivated by his ethical and safety concerns.

Why Jacob Coxon’s Resignation Matters

The controversy is unusual because Coxon’s warnings were partly validated by employees who remained inside Anthropic. Current researchers openly acknowledged that they believe advanced AI could pose catastrophic risks and that they do not yet have a complete solution for controlling superintelligence. [The Verge]

The disagreement is therefore not simply between people who believe AI is dangerous and companies that say it is safe. The deeper dispute is whether organizations that openly recognize catastrophic risks can responsibly continue racing to build more powerful systems before those safety problems are solved.

Coxon concluded that they cannot. His former colleagues at Anthropic have not necessarily agreed that resignation or a development pause is the correct response, but several publicly agreed that the underlying danger he described is real. [Axios]