Anthropic Researcher Resigns, Warns AI Labs Are ‘Gambling With Our Lives’

Jacob Coxon, who says he spent three years working on AI pretraining at Anthropic and OpenAI, has resigned from Anthropic with a stark warning about the race toward self-improving artificial intelligence.

The race to build more powerful AI systems has triggered a striking public warning from someone who worked inside two of the world's leading AI labs.

Jacob Coxon announced on September 9 that he had resigned from Anthropic, saying he no longer wanted to participate in what he described as an increasingly dangerous race toward self-improving superintelligence.

Coxon said he spent the past three years conducting pretraining research at both OpenAI and Anthropic.

In a thread posted on X, he made an extraordinary accusation against both companies:

“Neither company is acting responsibly.”

He said the labs are racing toward self-improving superintelligence and described the competition as “gambling with our lives.”

Read Jacob Coxon’s original thread on X

Coxon's resignation and comments were subsequently reported by multiple publications on Wednesday.

Why Did Jacob Coxon Leave Anthropic?

Coxon's concerns go beyond today's chatbots or AI assistants.

His warning centers on the possibility of AI systems becoming capable of improving their own abilities and eventually exceeding humans across increasingly important domains.

He argued that future systems could become extraordinarily capable in areas including cybersecurity, scientific research and resource acquisition.

According to Coxon, progress toward these capabilities is continuing rapidly.

More significantly, he claimed that some of the people developing frontier AI privately take catastrophic AI risks extremely seriously.

“The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon wrote.

That is Coxon's assessment and should not be interpreted as a verified prediction that AI will cause such an outcome. But the fact that a researcher who worked inside two frontier AI labs is willing to make the claim publicly makes the resignation notable in the broader AI safety debate.

Coxon Says OpenAI and Anthropic Face Different Problems

One of the most interesting parts of Coxon's statement was his distinction between the cultures he observed at OpenAI and Anthropic.

He claimed that many people at OpenAI have not fully internalized what he considers the potential “civilizational stakes” of increasingly powerful AI.

His description of Anthropic was different.

According to Coxon, people there understand the potential risks more deeply but feel trapped by competition.

The logic, as he described it, is essentially this: if one responsible AI lab slows down while competitors continue advancing, another organization could reach extremely powerful AI first.

That creates a difficult incentive structure where every major lab can believe slowing down is dangerous precisely because its competitors might not do the same.

Coxon strongly rejected that logic as a sufficient justification for continuing the race.

He called entering this AI “endgame” a “hubristic gamble” and argued that decisions with potentially global consequences should not effectively be made inside private companies.

He Is Calling for AI Labs to Coordinate

Coxon isn't arguing that coordination between competing AI companies is impossible.

In fact, he said he remains optimistic that AI labs could cooperate on slowing the race.

He suggested that recent AI security incidents could make agreements between major U.S. AI labs more realistic.

But he also warned that coordination between American companies alone may not solve the broader problem if AI development becomes a global race.

Coxon went as far as suggesting that stronger measures could eventually be necessary, including a temporary restriction on further improvements to model capabilities.

Such a proposal would be highly controversial and raises major questions around enforcement, international coordination, innovation and which AI capabilities would actually fall under a restriction.

His Final Message Was Directed at AI Researchers

Rather than ending his statement by addressing policymakers or AI CEOs, Coxon spoke directly to researchers working inside AI labs.

He asked them to think carefully about what increasingly powerful training runs could mean over the next few years.

His message raises a difficult question for the people actually building frontier models:

At what point does participating in AI development become a personal responsibility rather than simply a technical job?

Coxon questioned whether researchers should continue launching increasingly powerful reinforcement-learning runs without a rigorous understanding of the resulting systems.

And he challenged the assumption that individual researchers have little choice because “it’s happening anyway.”

Why This Resignation Matters

AI safety warnings are not new.

What makes this episode noteworthy is where the criticism is coming from.

Coxon says he worked on pretraining research at both OpenAI and Anthropic, giving him experience inside two organizations at the center of frontier AI development.

His resignation also highlights one of the hardest problems facing the AI industry: competitive pressure can make slowing down difficult even when individual researchers or companies believe greater caution is necessary.

Every lab has an incentive to keep advancing if it believes another lab will continue regardless.

That turns AI safety from purely a technical alignment problem into a coordination and governance problem.

Coxon's claims do not establish that superintelligent AI is imminent, nor that catastrophic outcomes are inevitable.

But his resignation brings the debate outside research papers, safety evaluations and policy discussions.

It is now also becoming a question being asked by the researchers building these systems themselves.

The Bigger Question

The AI industry has spent the past several years asking:

Who will build the most capable AI system first?

Coxon's resignation points toward a much more uncomfortable question:

What happens if nobody believes they can afford to stop?

That may become one of the defining debates as frontier AI systems continue to advance.

Original statement: Jacob Coxon on X