The AP News Anthropic AI extinction warning landed not from a boardroom briefing but from a park bench in San Francisco’s Alamo Square, where researcher Jacob Coxon posted his resignation on 8 September 2026, forfeiting equity that would have vested within two months.
That detail matters. People who quit comfortable, lucrative jobs two months before a financial windfall are not usually doing it for attention.
When Insiders Start Resigning Over Safety
Coxon had, according to Time, shifted his duties just a week before resigning (from training machine-learning models to safety research) but found the change did nothing to lift his ‘feeling of impending doom.’ His stated reason for leaving was specific: he did not want to participate in building AI systems capable of improving themselves, fearing such systems could spiral beyond human control.
‘Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives,’ he wrote, targeting both Anthropic and his former employer, OpenAI.
What followed was more striking than the resignation itself. Evan Hubinger, who describes himself as a lead in Anthropic’s alignment division, publicly backed Coxon’s assessment. ‘We really do earnestly believe AI could kill all humans!’ he wrote, adding: ‘I personally think it is >10% 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.’
That is a named, active senior employee assigning double-digit probability to human extinction. Samuel Marks, Anthropic’s ‘scalable oversight lead’, posted separately (stressing he spoke in a personal capacity) that ‘the more senior the employee, the more concerned they are.’
Anthropic’s corporate response was predictable. The company pointed to its Responsible Scaling Policy, first published in September 2023, which ties capability thresholds to mandatory upgrades in safety standards. It called itself a pioneer in ‘mechanistic interpretability’ and said it ‘aggressively’ tests models for dangerous capabilities in cybersecurity and biology. None of that directly answers why two of its own researchers are publicly assigning meaningful probability to extinction on their watch.
Anthropic was founded in 2021 by a group who left OpenAI specifically over safety concerns, and has consistently marketed itself as the more responsible of the frontier labs. That positioning now sits awkwardly alongside its own employees’ public statements.
The Anthropic AI Extinction Warning Congress Cannot Ignore
The political response arrived quickly, and was more substantive than the usual senatorial social media repost. Bernie Sanders introduced the Ban Artificial Superintelligence Act on 23 September 2026, co-sponsored by Representative Greg Casar (D-Texas). The bill would permanently ban the development and deployment of superintelligent AI systems and pause advanced AI development until a new Cabinet-level Department of Artificial Intelligence (led by a Senate-confirmed secretary) had established safety rules.
The penalties are designed to concentrate minds. Violations would carry what the sponsors call a ‘corporate death penalty’ plus up to 20 years in prison, modelled explicitly on penalties for unlawfully developing nuclear weapons. The bill defines ‘Artificial Superintelligence’ as any system capable of matching or exceeding human cognitive performance across a broad range of domains, or one capable of planning to undermine the US government, according to the official bill summary. At 19 pages, it is brief for something carrying that kind of ambition.
A number of current employees at leading AI companies signed a statement of support for the bill, reported first by the Associated Press. That is a meaningful signal: people still drawing salaries from the companies targeted by the legislation publicly endorsing a bill that would criminalise their employers’ core work.
The bill’s prospects in the Senate are uncertain. Senate Minority Leader Chuck Schumer indicated he would urge colleagues not to support it, backing a separate Democratic bill he says contains mandatory enforcement mechanisms rather than voluntary ones, according to Politico. Senate Commerce Committee Chair Ted Cruz is working on his own bipartisan bill with Majority Leader John Thune and Senator Amy Klobuchar, with no timeline for releasing text.
My read is this: the Anthropic AI extinction warning is not credible because Coxon quit, or because senators are paying attention. It is credible because the people still inside the building, still on the payroll, are saying the same thing privately, and occasionally, publicly. That is the data point that the industry’s reassurances cannot paper over. The real question is whether any of the competing bills in Washington will move fast enough to matter, or whether the race Coxon described will simply continue, slightly more loudly, while legislators argue over enforcement mechanisms.
