Two researchers who held positions at some of the world's most prominent AI labs have departed in recent months, each citing the same core concern: the companies building the most capable AI systems are moving faster than their ability to keep those systems safe. Their warnings, backed by colleagues still inside the labs, paint a picture of an industry where the people closest to the technology are increasingly alarmed by where it's headed.

Jacob Coxon, a 27-year-old pretraining researcher, resigned from Anthropic on September 9 after spending three years doing core model training work, first at OpenAI and then at Anthropic. His departure announcement on X reached more than 100 million people overnight. Alexander Turner, an AI safety research scientist, left Google DeepMind in June after more than two years, citing the company's decision to sign a Pentagon contract allowing its Gemini AI to be used for classified military operations without restrictions on autonomous weapons.

Neither researcher moved to a competitor. Both walked away from positions at the top of the field, and both say the industry's internal culture has shifted in ways that make safety concerns difficult to address from within.

The Coxon Resignation

Coxon's departure letter, posted to X on September 9, is blunt. "I resigned from Anthropic today. 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 research focused on pretraining, the phase where foundational AI models learn general capabilities before being fine-tuned for specific tasks. He worked on GPT-4o at OpenAI before moving to Anthropic. In interviews with the Wall Street Journal and NBC News, he described researchers inside frontier labs increasingly using words like "crunchtime" and "endgame" to describe the pace of capability development.

"The people building AI earnestly believe that it could kill us all by the end of the decade," Coxon wrote. "This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible, but I hear the same people express fear privately."

Coxon distinguished between OpenAI and Anthropic in terms of their internal cultures. At OpenAI, he said, "many have not deeply internalized the civilizational stakes." At Anthropic, the risks are well understood, but the company is "locked in a race to get there first" because "they believe no one else will act responsibly, so they must do it themselves, despite the risk."

He argued that the industry's standard justification for continuing to build more powerful systems, that someone else will do it anyway, is insufficient. "Accepting this race and entering the 'endgame' is a hubristic gamble that should not be launched from a private company's Slack," he wrote. "Attempting to speedrun alignment should require extraordinary confidence that there are no better trajectories available."

The Turner Departure

Turner's reasons for leaving Google DeepMind are different in specifics but overlap in principle. He worked on AI safety for more than two years and proposed a framework for military AI that would have required human control over AI targeting systems. He sent the proposal to DeepMind CEO Demis Hassabis, who directed him to have it evaluated by senior policy staff.

After some discussions, Turner stopped receiving responses. Shortly after, the Department of Defense confirmed it had signed a deal with Google. "At that point, I couldn't stay at Google in good conscience, so I left," Turner wrote in a blog post published in July.

Turner's departure reflects a different axis of the safety debate: not just whether AI systems might become uncontrollable, but whether the institutions deploying them can be trusted to impose meaningful constraints. The Pentagon contract allows Gemini to be used for any lawful government purpose, with no restrictions on autonomous weapons or surveillance applications.

In September, Turner publicly backed Coxon's warning, saying the concerns were familiar from his own work on advanced AI. His endorsement is notable because it connects two separate threads of the safety conversation, the race toward superintelligence and the militarization of AI systems, into a single picture of an industry moving in directions its own researchers find troubling.

The Internal Response

The most striking aspect of Coxon's departure is not that he left, but that his colleagues inside Anthropic publicly agreed with him. Evan Hubinger, who leads Anthropic's alignment stress-testing team, replied directly: "Jacob is correct here, we really do earnestly believe AI could kill all humans! I personally think it is greater than 10% within the next decade."

Hubinger added a caveat: "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to." He distinguished between risks from present models, which he described as low, and risks from future systems that the company is racing to build.

Samuel Marks, an Anthropic safety researcher speaking in a personal capacity, also backed Coxon's broader warning. The pattern is unusual in an industry where employees typically express disagreements privately, if at all. The public nature of these endorsements suggests that the safety concerns have moved beyond individual frustration into something closer to consensus among researchers who spend their days working on the exact problems that need solving.

The Earlier Warning

Coxon's departure is not the first time an Anthropic safety researcher has left with a public alarm. In February, Mrinank Sharma, who led the company's Safeguards Research Team, resigned with a letter stating that "the world is in peril" from "a whole series of interconnected crises."

Sharma's team had been established just over a year earlier to tackle AI security threats including model misuse, bioterrorism prevention, and catastrophe prevention. His accomplishments included developing defenses against AI-assisted bioweapons and researching how AI assistants could distort human behavior. He described "repeatedly" seeing how hard it is to let values govern actions, both within Anthropic and throughout broader society.

Instead of moving to another AI lab, Sharma announced plans to pursue a poetry degree in the UK and "become invisible for a period of time." The contrast between the urgency of his warning and the quietness of his response underscores the difficulty researchers face in finding ways to act on their concerns within the existing industry structure.

What the Companies Say

Neither Anthropic nor OpenAI has responded to detailed questions about the specific claims made by their former researchers. Both companies have said in recent weeks that they are taking safety seriously and slowing some development processes. Anthropic amended a key safety pledge this year, dropping a commitment not to train more powerful models without adequate safeguards and replacing it with safety roadmaps and risk reports.

Senator Bernie Sanders has announced plans to introduce a bill that would pause AI development and create an industry regulator. An upcoming meeting between American and Chinese officials on AI safety later this month represents another potential mechanism for coordination. More than 1,300 employees from frontier AI companies signed a July statement calling for deliberate mechanisms to pace the development of advanced AI systems.

But the researchers who have left say these measures lag behind the pace of capability development. Coxon's timeline is blunt: "We're on track for a lot of the most aggressive of these scenarios where by the end of next year things could be out of control already."

The question facing the industry is no longer whether these researchers believe the risks are real. They clearly do, and they're saying so publicly. The question is whether anyone with the power to change the trajectory is listening.