Something unusual has been happening inside the world's biggest AI companies. People who spent years building the technology are walking out the door, and instead of quietly moving on, several have gone public with warnings that sound less like a standard resignation letter and more like a caution flag being raised in real time.
This isn't one isolated event. It's a pattern that picked up speed fast in September 2026. Below is what's actually documented, attributed to the people and outlets who reported it, what's personal opinion rather than established fact, and where the real disagreement sits.
The resignation that started it: Jacob Coxon
On September 8, 2026, a 27-year-old researcher named Jacob Coxon resigned from Anthropic and posted his reasoning publicly on X. He had spent the previous three years doing pretraining research at both OpenAI and Anthropic, giving him direct experience inside two of the field's leading labs before he left.
His opening line set the tone: "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."
Axios, which conducted a follow-up interview with Coxon, reported that his post drew more than 115 million views on X within days. IBTimes, citing the same X thread, put the figure above 100 million shortly after it went live, and later coverage from Startup Fortune cited view counts exceeding 150 million as the story continued to spread. In that Axios interview, Coxon said "if you're under pressure to race, you have to cut corners" or skip steps in the oversight process, though he also told Axios he personally hadn't witnessed Anthropic compromising safety to outpace competitors. He added nuance rather than pure alarm, telling Axios the industry sometimes shows what he called "excessive paranoia" about OpenAI and China, which he said can itself be used to justify pushing ahead regardless.
One detail made his exit harder to dismiss as a publicity move. Axios reported that Coxon left Anthropic after four months there, roughly two months before his equity was due to begin vesting, and that he forfeited that entire unvested stake by leaving when he did. "I no longer have anything to gain by juicing up Anthropic's valuation," Coxon told Axios. He still holds equity from his earlier time at OpenAI, according to Axios, so the claim isn't that he has zero financial stake in the industry's fortunes. But he did walk away from the specific compensation tied to the company he was resigning from and publicly criticizing.
Some commentators on social media, including some conservative and tech-industry figures, suggested the resignation looked staged to build political support for AI regulation. Reporting that covered these claims noted that critics offered little direct evidence to back that theory up. That's worth stating plainly: it's an accusation that circulated online, not a documented fact, and it should be read as such.
Why Evan Hubinger's response matters
Coxon wasn't left to make this case alone. Evan Hubinger, who leads alignment science research at Anthropic and remains employed there, replied publicly in support of Coxon's warning: "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. 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."
This detail matters for a specific reason. Hubinger isn't a disgruntled former employee speaking from the outside. He's a senior safety researcher still working at the company being criticized, publicly agreeing with a departing colleague's warning while remaining on staff.
It's important to be precise about what his number actually represents. Hubinger's "greater than 10%" figure is his own personal estimate, stated in his own words on his own account, not a peer-reviewed calculation or an official company forecast. Anthropic has separately published its own risk assessment materials, which acknowledge this category of concern while describing the current probability of AI actually gaining that kind of uncontrolled power as low. Both statements, a senior researcher's personal double-digit estimate and the company's more measured official position, are on the record at the same time, and they don't fully line up with each other.
A third researcher, Sam Marks, added his own account around the same period, writing publicly that staff keep building the technology despite believing in genuine extinction-level risk, partly due to financial incentive and partly out of fear that a less cautious competitor will build it anyway if they don't.
This isn't the first AI safety exodus, but it's the loudest one yet
None of this is entirely without precedent. Geoffrey Hinton, often called the godfather of AI, left a senior position at Google back in 2023 specifically so he could speak freely about these risks without a company's public relations concerns shaping what he was allowed to say. Anthropic itself exists as a company because a group of former OpenAI employees left that company over disagreements about how safety was being handled internally.
What's different in 2026 is the volume and the visibility. NBC News reported that more than a dozen AI company employees told its reporters they're genuinely frightened by the direction of the industry, even without being able to name the single failure mode they're most worried about. Within roughly ten days of Coxon's resignation, at least two more safety researchers left major labs with comparable public warnings attached to their departures. Around the same period, executives from companies that normally compete fiercely, including OpenAI, Anthropic, and xAI, separately acknowledged in public statements that the industry may need to slow down. OpenAI's chief scientist, Jakub Pachocki, wrote in a blog post, later reshared by CEO Sam Altman, that no one is prepared for the consequences of a continued rapid rise in machine intelligence.
Why researchers keep building this anyway
The recurring explanation across these accounts is a race dynamic. Researchers describe believing the risk is real while feeling structurally unable to stop it alone, because unilaterally slowing down doesn't reduce the actual risk if competitors keep building regardless. Coxon's own comments reflect this logic. He told Axios he doesn't believe mass resignation is the right response, and suggested that people who share his concerns are often better positioned pushing for change from inside these companies rather than leaving them entirely.
The regulatory capture argument, and who's actually making it
Here the story splits into genuinely contested territory, and both sides deserve to be named rather than summarized as vague "some critics say" framing.
Critics who question the convenience, if not the sincerity, of these warnings call this regulatory capture. Their argument runs like this: loud public warnings about extinction-level risk help build political support for expensive safety regulations, regulations that only large, well-capitalized labs can realistically afford to comply with, while smaller startups and open-source competitors get squeezed out. David Sacks, who has advised the Trump administration on AI and crypto policy, wrote publicly that he sees Anthropic specifically as running what he called a sophisticated regulatory capture strategy built on fear.
A more specific version of this argument connects the warnings to geopolitical competition. Some commentators writing in outlets like the Globe and Mail have argued that highlighting existential-risk claims could help build political support for restricting Chinese AI models, which would shield US labs from cheaper foreign competition. That specific connection is an interpretation offered by named commentators in opinion and analysis pieces. It hasn't been confirmed as an actual, stated motive by anyone at the companies involved, and it remains genuinely contested.
Researcher Meredith Whittaker offered a sharper framing that critics have revived in this context, describing existential-risk warnings as "ghost stories" that function as advertisements for a technology only a handful of companies currently control.
What critics of the doom narrative actually argue back
On the other side, economist Tyler Cowen pushed back directly on the marketing-stunt framing in a widely shared piece, pointing out that Sam Altman, Dario Amodei, and Elon Musk were all talking publicly about AI risk well before any of them ran a company positioned to financially benefit from that conversation. Cowen argued that treating repeated, consistent statements from people with no product to sell, like Hinton, as a coordinated PR campaign requires ignoring a fair amount of contrary evidence.
Coxon tried to head off this exact criticism in his own resignation post, writing directly: "This is not a marketing stunt." Kat Duffy, a senior fellow for digital and cyberspace policy at the Council on Foreign Relations, told Foreign Policy she doesn't agree with the theory that these warnings are necessarily PR moves timed around anticipated IPOs from OpenAI and Anthropic.
Can both explanations be true at once?
A growing number of policy analysts argue yes, and this is probably the most useful way to make sense of the disagreement rather than picking a side outright.
There are at least three distinct possibilities worth separating. First, individual researchers can genuinely and personally fear catastrophic outcomes from their own work, independent of any company's business interests. Second, companies themselves can hold real internal safety concerns while still pursuing aggressive commercial deployment, since internal caution and external competitive pressure don't automatically cancel each other out. Third, public warnings about risk can create regulatory and competitive advantages for the companies making them, whether or not that outcome was ever the original intent behind the statements.
None of these three possibilities require the others to be false. A warning doesn't have to be manufactured to also happen to benefit the company or person voicing it. That combination, sincere personal concern that coincides with convenient business outcomes, is probably closer to what's actually happening than either a pure "they're all lying" reading or a pure "it's all pure altruism" reading.
What we actually know, and what we don't
| Claim | What the evidence shows |
|---|---|
| Researchers have resigned citing AI safety concerns | Confirmed. Multiple named individuals, on the record, with public statements verified by Axios, NBC News, and other outlets. |
| Jacob Coxon forfeited unvested Anthropic equity to leave | Confirmed by Axios, corroborated in follow-up reporting from IBTimes and others. |
| Some researchers personally estimate greater than 10% extinction risk within a decade | Confirmed as a personal estimate from named individuals like Evan Hubinger. Not an industry-wide consensus figure. |
| All AI researchers believe extinction risk is high | Not established. Public estimates range from near 0% to over 20% depending on who is asked. |
| The warnings are proven regulatory capture | Not established. This is a contested interpretation argued by named critics such as David Sacks, not a confirmed or admitted motive. |
| A majority of Americans see at least moderate AI risk | Confirmed by a September 2026 POLITICO/Public First poll of 2,064 US adults. |
| AI will definitely cause human extinction | Not established by any source cited in this article or elsewhere. |
What the public actually thinks
This conversation has moved well past Silicon Valley. POLITICO reported that a survey conducted by Public First, covering 2,064 US adults between September 13 and 15, 2026, with a margin of error of roughly 2.2 percentage points, found that 63% of respondents saw at least a moderate risk that advanced AI could eventually destroy humanity. That combined figure blends three separate response categories: 17% called the outcome almost certain, 20% called it a significant risk, and 26% called it a moderate risk, so the widely reported headline number reflects a range of conviction levels rather than one uniform belief. The same survey found 48% of respondents favored slowing advanced AI development, compared with 31% who favored continuing at the current pace.
Among researchers actually working on this technology day to day, there's far less agreement. Estimates of catastrophic AI risk, often shortened to "P(doom)" in industry conversation, vary enormously depending on who's asked. Company leaders and senior safety researchers, including Dario Amodei and Evan Hubinger, have cited figures roughly in the 10% to 25% range. Aggregated forecasts pulled from prediction platforms like Metaculus and Manifold Markets tend to cluster closer to 5% to 15%. Researchers focused specifically on today's systems, such as Meta's Yann LeCun, put the number near zero, arguing the extinction framing distracts attention from AI harms that are already measurable right now, including algorithmic bias, surveillance, and misinformation.
The bottom line
A wave of AI researchers walking away from major labs while warning their own work could end badly isn't something that happens often, and it isn't something to dismiss without a closer look. It also isn't proof of anything on its own. Some well-informed people in this field genuinely believe the danger is real and near-term. Others, just as informed, believe the public framing of that danger is doing convenient work for the companies behind it. Both readings deserve to be taken seriously on their own terms. Right now, nobody outside these labs, and arguably not everyone inside them either, can say with confidence which one is closer to the truth.
Sources & Further Reading
- Axios — Anthropic whistleblower gave up his equity to leave the company
- Free Press Journal — Jacob Coxon says he quit before equity vested over AI safety fears
- Free Press Journal — Meet Jacob Coxon: profile and industry reaction
- IBTimes — Reporting on Coxon's resignation post and equity decision
- Startup Fortune — Evan Hubinger's public statement and Washington's reaction
- Oliver Willis — Explainer addressing claims the resignation was staged
- NBC News — AI insiders describe fear about the industry's direction
- Foreign Policy — Analysis of motives behind AI existential-risk warnings, including Kat Duffy's comments
- Marginal Revolution — Tyler Cowen's case against the regulatory-capture reading
- Remio — Detailed breakdown of the POLITICO/Public First September 2026 poll
FAQs
Why did Jacob Coxon leave Anthropic? Coxon resigned on September 8, 2026, saying he believed Anthropic and OpenAI, his two previous employers, were racing toward increasingly powerful AI systems faster than they could safely manage. He told Axios he left before his equity vested, which he pointed to as evidence he had little personal financial reason to raise the alarm.
Who is Jacob Coxon? Coxon is a 27-year-old AI researcher who spent roughly three years doing pretraining research at OpenAI and then Anthropic before resigning from Anthropic after about four months there. His resignation post drew more than 100 million views on X, according to Axios and IBTimes.
What did Evan Hubinger say about AI extinction risk? Hubinger, who leads alignment science research at Anthropic and remains employed there, publicly stated he personally believes there is more than a 10% chance AI could cause human extinction within the next decade. This is his personal estimate, not an official company forecast or a scientific consensus figure, and Anthropic's own published risk materials describe current uncontrolled-power risk more cautiously.
Are AI researchers really quitting because of safety concerns, or is something else going on? Both explanations have real support. Some researchers, like Coxon and Geoffrey Hinton before him, appear to have left specifically over safety concerns and at real personal cost, including forfeited equity. Critics, including former Trump AI advisor David Sacks, argue the broader wave of public warnings also conveniently supports regulations that would favor large, established labs. Analysts increasingly suggest both dynamics can be happening at the same time.
What is P(doom)? P(doom) is informal industry shorthand for the estimated probability that AI causes a catastrophic or extinction-level outcome for humanity. Estimates vary widely, from near 0% among researchers focused on current systems, like Yann LeCun, to over 20% among some industry leaders and safety researchers.
How likely is AI to cause human extinction, according to experts? There's no consensus. Estimates cited by industry leaders and safety researchers, including Dario Amodei and Evan Hubinger, generally fall between 10% and 25%. Aggregated prediction-market forecasts tend to cluster closer to 5% to 15%. Some prominent researchers estimate the risk at close to zero.
Why are AI companies continuing development despite safety concerns? Researchers who've spoken publicly, including Coxon, point to competitive pressure as the main reason. The concern is that if one company slows down unilaterally, a less cautious competitor will keep building anyway, so the perceived risk of falling behind outweighs the incentive to pause alone.
What is AI regulatory capture? It's the theory that AI companies' public warnings about existential risk help build political support for safety regulations that only large, well-funded labs can afford to comply with, effectively disadvantaging smaller competitors and open-source developers. It's a contested interpretation, argued publicly by named critics including David Sacks, rather than a confirmed or admitted motive.
Do AI researchers agree with each other about extinction risk? No. Estimates range widely even among people working at the same company. Anthropic alone has had a senior researcher, Evan Hubinger, cite a greater than 10% personal estimate while the company's own published risk materials describe current uncontrolled-power risk as low, showing the disagreement exists even internally.
What are the immediate risks of AI, besides extinction scenarios? Researchers skeptical of the extinction framing, including Yann LeCun, point to harms that are already measurable today, including algorithmic bias, surveillance misuse, deepfakes and misinformation, and job displacement. Critics of the doom narrative argue these near-term issues deserve more regulatory attention than they currently receive.
Did any AI company executives respond to these resignations? Yes. OpenAI's chief scientist, Jakub Pachocki, wrote publicly that no one is prepared for the consequences of a continued rapid rise in machine intelligence, a post later reshared by CEO Sam Altman. Executives at OpenAI, Anthropic, and xAI separately acknowledged around the same period that the industry may need to slow down, an unusual moment of public agreement among companies that normally compete fiercely.
Written by Muntazir Mahdi, founder of ANFA Technology, for AI Future Insights. AI Future Insights covers artificial intelligence, automation, and future tech for readers who want signal over hype, built and maintained by a Karachi-based team working on privacy-first software including Canvas Convert Pro. This piece is sourced from reporting by Axios, NBC News, IBTimes, Free Press Journal, Startup Fortune, and Foreign Policy on the September 2026 wave of AI safety resignations, commentary from Marginal Revolution and the Globe and Mail on the regulatory capture debate, and polling data from POLITICO and Public First on public attitudes toward AI risk.