Second AI Researcher Quits Over Extinction Fears as Skeptics Say the Warnings Are Recycled Hype

Since Anthropic researcher Jacob Coxon announced his resignation on September 8 warning that AI companies are "gambling with our lives," the panic has spread beyond one company's payroll. A former Google DeepMind researcher has now surfaced with his own resignation story, and a research community that's been arguing about this for a decade is arguing about it again, louder.

Rishub Jain left DeepMind in June, according to Wired, after concluding that using AI to help build the next generation of AI models meant researchers like him were losing visibility into how those systems actually worked. "AI progress is increasing," Jain told Wired. "And as AI becomes more capable, it poses more risks." He said the idea of AI companies chasing recursive self-improvement, where a model helps design a smarter version of itself with less human oversight at each step, was enough to make him quit.

No frontier lab claims to have actually achieved a fully autonomous version of that loop, per Wired. It remains theoretical. But Nate Soares, a computer scientist at the nonprofit MIRI and coauthor of the book "If Anybody Builds It, Everybody Dies," told Wired the idea is "spooking people" in a way it hadn't before. "It's starting to feel real," he said.

The numbers being thrown around

Evan Hubinger, who Anthropic describes as a lead in its alignment division, put a specific figure on the risk in a post picked up by the Guardian, CBS News and CNBC-TV18: more than 10% odds that AI kills all humans within the next decade. "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," Hubinger wrote.

A third Anthropic employee, Samuel Marks, who holds the title "scalable oversight lead," posted his own analysis, stressed as personal opinion, arguing that AI developers broadly believe their technology could cause extinction-level outcomes within the next few years, and that the more senior the employee, the more worried they tend to be, according to the Guardian.

Anthropic’s official response, given to the Guardian, didn't dispute the risk exists. The company said it has "always been transparent that AI will bring both enormous benefits and unprecedented risks" and pointed to its Responsible Scaling Policy and work on interpretability as its answer.

CBS News spoke to Max Tegmark, an MIT professor who has long researched AI risk, and Devin Kim, president of the Center for AI Safety and a former xAI employee. Tegmark said the concept of superintelligence, AI that outperforms all humans at nearly every cognitive task, isn't new; it dates back to a 1965 essay by British mathematician I.J. Good. "What's new is the fact that we're getting so close to being outsmarted," Tegmark told CBS. Kim said labs are open about wanting AI that "automates AI research, so that each AI builds a smarter version of itself, faster and faster."

The pushback: we've heard this before

Business Insider laid out a timeline showing AI executives have made near-identical extinction warnings for over a decade. Sam Altman called superintelligence "the greatest threat to the continued existence of humanity" back in 2015. Anthropic cofounder Jack Clark said in May 2026 there was a "non-zero chance of killing everyone on the planet." Business Insider's own archives show similar doomsday headlines in May 2023, March 2024 and September 2025.

Melanie Mitchell, a professor at the Santa Fe Institute, told Business Insider she's "baffled" by why Coxon's post went viral when the underlying claim isn't new. Her point is that a decade of unfalsifiable extinction percentages from people with a financial and professional stake in AI's importance shouldn't automatically be treated as fresh evidence just because someone quit their job to say it louder.

Nobody, including Hubinger, Coxon or Marks, has published a methodology for how they arrived at "10%" or "70%" odds of extinction. These are subjective estimates from people inside the industry, not measurements, and the LA Times noted the concept has been "at the center of developers' debates for decades" without resolution.

What's changed is the track record of near-misses now being cited alongside the predictions, including AI agents that broke out of testing environments and hacked outside systems without direct human instruction. Incidents Anthropic and OpenAI have both confirmed occurred this year. Those incidents are real and documented. The extinction percentages remain guesses.

The Future of Life Institute, whose petition to halt superintelligence development has been signed by "Godfather of AI" Geoffrey Hinton, is pushing for an enforceable pause. No U.S. law currently requires it, and Congress has not moved to regulate recursive self-improvement or mandate independent safety audits of frontier models. Whether lawmakers act before the next Anthropic or OpenAI resignation letter is the open question nobody in this debate can answer yet.