Laboratory of the Future

Dr. Ricardo Petrissans Aguilar.

Some resignations are an administrative formality, and some resignations are an earthquake. Jacob Coxon’s undoubtedly belongs to the second category.

On Tuesday, September 8, at 6:47 in the morning Pacific time, this 27-year-old British engineer opened his laptop in his San Francisco apartment, wrote for twenty minutes, and posted a thread on X that, within hours, surpassed 58 million views. Before noon, his name was a trending topic in fourteen countries. Before the day was over, he had formally resigned.

An engineer who knew both OpenAI and Anthropic

Coxon is not an outside critic. He is not an armchair philosopher or an alarmist columnist who has never written a line of code. He worked for three years at OpenAI, where he took part in the development of GPT-4 and GPT-4.5, and then joined Anthropic —the company founded by former OpenAI employees that promised to prioritize safety over speed— precisely because he believed it was the right place. «I resigned from OpenAI to join Anthropic because I initially considered it more cautious», he wrote. It was his personal commitment to doing things right.

And now he was leaving both.

«The people creating artificial intelligence sincerely believe that these systems could kill us all by the end of this decade», he wrote. «No other human activity poses this level of threat».

The phrase most often repeated in headlines around the world was another: «They are gambling with our lives». But Coxon’s thread was more nuanced and, for that reason, more devastating. He was not accusing specific individuals of malice. He was describing a system —a competitive race between companies— that pushes all its participants forward even when they know the precipice is close.

The trap of getting there first

What Coxon revealed was a crucial distinction. According to his testimony, there is a difference between OpenAI and Anthropic. The former «may not have deeply internalized the civilizational challenges», while the latter «understand them well». But both are «trapped in the race to be first», even when, «in the most optimistic case», it is a matter of «getting there first to act responsibly».

That phrase —«getting there first to act responsibly»— sums up the trap the entire industry is in. Nobody wants to be the one who puts on the brakes. Everyone fears that, if they stop, someone else will keep going. And within that racing logic, safety becomes a luxury nobody can afford.

Coxon sees no way out other than coordination among the different AI projects and the companies involved. He feels that this is not happening. That is why he advocates costly actions, such as a temporary ban on improving model capabilities. And he ends his thread with a question aimed directly at his colleagues: «Do you want to enable superintelligent reinforcement learning without a rigorous understanding of its mind? Are you going to bow your heads because “after all, it is already happening”, or take this moment to advocate for different conditions?».

Evan Hubinger's backing

What turned Coxon’s resignation into a turning point was not just his testimony. It was what happened afterward. Evan Hubinger, head of Alignment Science at Anthropic —that is, the person whose job is precisely to work out how to make AI systems safe— did not refute Coxon. He backed him.

«We wholeheartedly believe that AI could kill all humans», Hubinger wrote. «Personally, I think it is more than 10% within the next decade». And he added something even more disturbing: Anthropic, one of the world’s leading companies in AI safety, «still does not have a plan to solve alignment for superintelligence, and we are not clearly on track to do so».

That «not yet» resonates like a verdict. Because if the technology’s own creators acknowledge that they do not know how to control it, what is left for the rest of us?

Hubinger clarified that his concern is not «current models», but «the superintelligence that emerges from recursive self-improvement». In other words, a system capable of redesigning itself to become more intelligent, which in turn redesigns itself to become even more intelligent, in a loop that accelerates exponentially. Such a system, if it is not perfectly aligned with human values, could pursue goals we do not understand, using methods we cannot foresee.

A chain of warning signs inside Anthropic

Coxon’s resignation was not an isolated act. In February 2026, Anthropic had already removed from its safety charter the commitment to suspend the development of its models if it could not control the risks. In June, the company suspended and temporarily deactivated its most powerful models, Mythos 5 and Fable 5, to review its digital security systems after a series of cybersecurity incidents that affected several versions of Claude. Another employee, Mrinank Sharma, had resigned in February.

What Coxon did was break the silence. Put names to what many in the industry were whispering in private. And in doing so, he forced Anthropic to publicly confirm what until then had only been hinted at: that the creators of the most powerful technology in history are afraid of what they are building.

Risk assessment