An Anthropic researcher’s resignation post racked up nearly 76 million views on X in under 24 hours this week, and by the time the dust settled, one of the company’s own alignment leads had publicly backed up his warning. Jacob Coxon, who spent the past three years doing pretraining research at both OpenAI and Anthropic, announced on September 9, 2026 that he had quit Anthropic, telling followers that “neither company is acting responsibly” and that both labs “are racing straight to self-improving superintelligence and gambling with our lives.” Within hours, Evan Hubinger, Anthropic’s alignment stress-testing team lead, replied on the same platform to say he agreed, and that he personally estimates the odds of AI-driven human extinction at “>10% within the next decade.”
The story moved fast even by AI-news standards. TechCrunch, Forbes, CNN Business, CNBC, Newsweek and Deadline all ran the story within a single news cycle, and the phrase “gambling with our lives” became a trending search term across the US and Europe by the morning of September 10. What makes this different from the usual AI safety op-ed is the source: Coxon is not an outside critic. He is a pretraining researcher who built models at both of the labs he is now criticizing, and the researcher who publicly agreed with him, Hubinger, still works inside Anthropic’s alignment team as of this writing.
What Jacob Coxon Actually Said
Coxon’s resignation arrived as a multi-part thread on X, not a formal statement or a company press release. He opened with a direct line: “I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic.” From there he moved to his central complaint, one that reporters immediately picked up as the story’s headline. Both companies, he wrote, are “racing straight to self-improving superintelligence and gambling with our lives.”
What separates Coxon’s post from a routine safety warning is his framing of intent. He is not accusing Anthropic or OpenAI of recklessness born of ignorance. He argued that the people building these systems know the stakes: “The people building AI earnestly believe that it could kill us all by the end of the decade.” That line reframes the entire debate. It is one thing to say a technology carries risk that its builders haven’t fully grasped. It is another to say the builders grasp it and are proceeding anyway because of competitive pressure, including rivalry with Chinese AI labs racing toward the same capability threshold.
Coxon’s resume gives the claim weight it wouldn’t carry from an academic or a policy commentator. Pretraining research sits at the center of how large models are built: it covers the data pipelines, scaling laws and compute allocation decisions that determine how capable a model becomes and how fast. Someone who has done that work at two of the three labs generally considered to be at the frontier (the third being Google DeepMind) has a vantage point most outside critics don’t have.
Anthropic’s Alignment Lead Doesn’t Push Back — He Agrees
The story could have ended as a single disgruntled-employee post that companies quietly wait out. It didn’t, because Evan Hubinger, who leads Anthropic’s alignment stress-testing team, responded publicly and did not dispute Coxon’s framing. His reply, in part: “I personally think it is >10% within the next decade.” A double-digit probability of an extinction-level outcome, stated on the record by a safety lead at the company building the models, is not a hedge. It’s a number that most companies would rather their own staff not put in writing on a public platform.
Hubinger’s reply also included an admission that is arguably more consequential for the industry than the headline risk percentage: Anthropic, he said, is trying its best but does not yet have a plan to solve alignment for superintelligence and is not clearly on track to do so. That is a distinction worth sitting with. Anthropic has spent three years positioning itself as the safety-first lab, the one founded specifically because its co-founders thought OpenAI was moving too fast. Having its own alignment lead say publicly that there’s no clear plan for superintelligence alignment undercuts that positioning in a way that no competitor’s PR team could manufacture.
None of this happened in a leaked memo or a whistleblower lawsuit. It happened in public, on X, between two people who both build the technology in question, in full view of investors, regulators and the companies’ own customers. That’s a break from the usual pattern of AI safety disputes, which tend to surface through anonymous sourcing or after-the-fact reporting.
Why This Resignation Is Different From Past AI Safety Departures
AI safety researchers have left frontier labs before, and several have gone public with concerns. Anthropic itself was founded in 2021 by former OpenAI staff, including Dario Amodei and Daniela Amodei, who left specifically over disagreements about safety and commercial pace. Since then, both OpenAI and Anthropic have seen individual safety and superalignment staff depart with public statements about insufficient investment in safety work.
The Coxon case differs in three ways. First, he worked pretraining, not the alignment or policy teams that have historically produced the more vocal departures — his complaint isn’t about his own team being under-resourced, it’s about the trajectory of the core capability work he was personally part of. Second, he named both labs he worked at rather than criticizing a single employer, which makes the story harder for either company to frame as a personnel dispute unique to their culture. Third, and most unusually, a sitting employee at one of the companies named in the resignation post responded by agreeing with the numbers rather than issuing a corporate rebuttal.
That third point is the one moving markets and prompting outlet after outlet to treat this as more than a viral tweet. Corporate crisis response typically calls for distancing a company from an ex-employee’s claims. Hubinger’s reply did the opposite.
Market and Industry Reaction
The immediate market reaction has been muted compared to the social media response, which is itself notable. Neither Anthropic nor OpenAI is publicly traded, so there’s no stock ticker to move the way Nvidia’s or AMD’s would on a supply-chain headline. But both companies carry enormous private valuations built substantially on the premise that they are managing frontier AI risk responsibly enough to keep operating with minimal regulatory intervention. A public statement from an internal alignment lead conceding no clear superintelligence safety plan exists is the kind of detail that due-diligence teams at every major investor and enterprise customer will now be asking about directly.
Enterprise buyers evaluating Claude or GPT-family models for regulated industries, healthcare and finance, tend to run vendor risk assessments that include AI safety governance questions. Those assessments now have a fresh, on-the-record data point to cite, and it didn’t come from a critic. It came from the company’s own alignment team.
Policymakers in Washington and Brussels have also seized on the timing. AI safety legislation has moved slowly in the US relative to the EU’s AI Act, and advocates for a federal AI safety body have repeatedly struggled to get traction against arguments that safety concerns are overstated or speculative. A frontier-lab alignment lead putting a specific number on extinction risk removes some of that “speculative” framing from the debate, whether or not regulators act on it quickly.
Anthropic and OpenAI: A Side-by-Side Look at the Safety Question
Coxon named both companies in his resignation post, but the two labs have taken different public postures on safety historically, which is part of why his blanket criticism landed as news. The table below lays out how each company has publicly positioned itself, based on their own stated commitments and structures as of September 2026.
| Factor | Anthropic | OpenAI |
|---|---|---|
| Founding premise | Founded in 2021 by ex-OpenAI staff citing safety-pace concerns | Founded in 2015 with an original nonprofit safety mission |
| Public safety framework | Responsible Scaling Policy, dedicated alignment stress-testing team | Preparedness Framework, safety and security committee |
| Named in Coxon’s post | Yes — his most recent employer | Yes — his prior employer |
| Internal response to the story | Alignment lead Evan Hubinger publicly agreed with the risk framing | No public on-the-record staff response reported as of Sept. 10, 2026 |
| Stated position on alignment readiness for superintelligence | Hubinger says no clear plan yet in place, per his own public statement | Not addressed in this story |
The asymmetry in the last two rows is the detail worth watching. Anthropic is the company that now has a public, named, on-the-record statement from a safety lead acknowledging the gap between current alignment work and what superintelligence would require. OpenAI has not offered a comparable public response tied to this specific story as of this writing, which leaves an open question about how it will respond as reporters continue to ask that company for comment.
A Brief History of AI Extinction-Risk Estimates
Numeric estimates of AI existential risk aren’t new to 2026. Researchers and industry figures have floated probability estimates in public forums and surveys for several years, often in the 5% to 20% range for catastrophic outcomes this century, sourced from AI safety researcher surveys and public statements by lab leadership. What’s different about the Coxon-Hubinger exchange isn’t the size of the number, it’s who said it and where. Survey respondents in academic polls are often anonymized. Lab leadership statements about long-term risk tend to arrive in carefully worded blog posts or congressional testimony, filtered through legal and communications review.
Hubinger’s reply had none of that filtering. It was a real-time public reply on a social platform, from a named employee still on staff, agreeing with an ex-colleague’s public resignation statement. That format, unscripted and immediate, is part of why the story spread the way it did. It read less like a corporate risk disclosure and more like two people who build these systems saying, in plain language, what they actually believe about the odds.
The Competitive Pressure Argument
Central to Coxon’s critique is a structural argument rather than a technical one: safety trade-offs become close to inevitable when multiple labs are racing toward the same capability milestones, especially when US labs are also framing the contest as a race against Chinese AI developers. That framing has shown up repeatedly in public comments from lab executives over the past two years, often used to justify faster release cycles and larger training runs. Coxon’s post flips that justification into an indictment: if competitive urgency is the reason safety work can’t keep pace, then the industry has built an incentive structure where no single company can unilaterally slow down without ceding ground to a rival that won’t.
That’s not a new observation in AI policy circles, but it usually comes from academics or advocacy groups pointing at the industry from outside. Hearing a version of the same argument from someone who spent three years inside two of the companies in question changes how much weight it carries in boardrooms and in regulatory hearings.
How the Story Spread: A Timeline
| Time / Date | Event |
|---|---|
| September 8, 2026 | Coxon leaves Anthropic, per subsequent reporting |
| September 9, 2026 | Coxon posts multi-part resignation thread on X |
| September 9, 2026 (same day) | Evan Hubinger publicly replies, agreeing with the risk framing and citing his own >10% estimate |
| September 9, 2026 (overnight) | Coxon’s thread approaches 76 million views |
| September 9-10, 2026 | TechCrunch, Forbes, CNN Business, CNBC, Newsweek and Deadline publish coverage |
| September 10, 2026 | Story becomes a top trending technology search topic in the US |
The speed of that spread mirrors past viral AI safety moments, but the view count on Coxon’s original thread stands out even against that history. A resignation post reaching tens of millions of views in under a day puts real pressure on both companies’ communications teams to respond in a way that satisfies reporters, employees and enterprise customers simultaneously, three audiences that often want different things from the same statement.
What Anthropic and OpenAI Are Likely to Do Next
Neither company had issued a formal corporate statement addressing the resignation directly as of this writing, beyond Hubinger’s personal reply. That gap itself is a signal. A company that wanted to shut the story down quickly would typically issue a statement reaffirming its safety commitments within hours. The absence of one, more than a day after the story broke, suggests internal deliberation about how to respond to an employee who effectively agreed with the departing researcher’s most serious claim.
Expect both companies to face direct questions from reporters and from enterprise customers about their internal safety timelines in the coming weeks. Given Hubinger’s specific point about lacking a clear alignment plan for superintelligence, Anthropic in particular is likely to face pressure to publish more detail about what its Responsible Scaling Policy actually commits the company to when models cross defined capability thresholds.
Predictions: Where This Story Goes From Here
- Anthropic will likely issue a more detailed public statement or blog post addressing Hubinger’s comments within the next one to two weeks, given the volume of press inquiries the story has generated.
- OpenAI will face growing pressure to produce an on-the-record response from its own safety leadership, given that Coxon’s criticism named the company directly.
- Expect renewed congressional and EU regulatory interest in AI safety disclosure requirements, using this story as a talking point even if no new legislation moves quickly.
- Other current and former frontier-lab researchers are likely to weigh in publicly in the coming days, either supporting or pushing back on Coxon and Hubinger’s framing.
- Enterprise AI procurement teams in regulated sectors will start asking vendors for more specific answers about internal alignment timelines during vendor risk reviews, using this story as precedent for why the question matters.
What This Means for Everyday AI Users and Developers
For developers building on Claude or GPT-family APIs, this story doesn’t change what’s shipping this week. Model behavior, pricing and rate limits remain governed by each company’s existing terms of service. What it does change is the context in which safety-related product decisions, like refusal behavior, red-teaming disclosures, or model card transparency, will be read going forward. A company whose own alignment lead has publicly stated there’s no clear plan for superintelligence safety will face more scrutiny the next time it ships a capability jump, not less.
For technology teams evaluating which AI vendor to build on long-term, the practical takeaway is to treat public safety statements from lab staff, not just official corporate blog posts, as material information worth tracking. This story is a reminder that individual researchers at these companies are willing to speak candidly in public, and that those comments can move faster and carry more weight than a quarterly safety report.
Frequently Asked Questions
Who is Jacob Coxon?
Jacob Coxon is an AI researcher who worked in pretraining research at both OpenAI and Anthropic over roughly the past three years. He announced his resignation from Anthropic in a public post on X on September 9, 2026.
Why did Jacob Coxon resign from Anthropic?
According to his own public statement, Coxon resigned because he believes neither Anthropic nor OpenAI is acting responsibly, and that both companies are racing toward self-improving superintelligence in a way he described as “gambling with our lives.”
Who is Evan Hubinger?
Evan Hubinger is Anthropic’s alignment stress-testing team lead. He publicly responded to Coxon’s resignation post, agreeing with the risk framing and stating he personally estimates the chance of AI causing human extinction at more than 10% within the next decade.
Did OpenAI respond to the resignation post?
As of this writing, no OpenAI staff member has issued a comparable public on-the-record response to Coxon’s post, even though he named OpenAI alongside Anthropic in his criticism.
What does “self-improving superintelligence” mean in this context?
It refers to AI systems capable of improving their own capabilities with limited human oversight, a scenario that AI safety researchers have long flagged as high-risk because it could accelerate capability gains faster than safety and alignment work can keep pace.
Has an Anthropic or OpenAI safety researcher put a number on AI extinction risk before?
Numeric risk estimates from AI researchers have appeared in surveys and public commentary before, but Hubinger’s reply is notable for being an immediate, public, on-the-record statement from a current safety lead at one of the companies actually building frontier models.
Is this expected to affect Anthropic’s or OpenAI’s business?
Neither company is publicly traded, so there’s no direct stock impact to track. However, both rely heavily on enterprise trust and investor confidence in their safety governance, and this story gives due-diligence teams a concrete, sourced data point to raise in vendor and investment reviews.
What should developers using Claude or GPT models do differently right now?
Nothing changes technically in the short term. It’s worth watching for any follow-up statements from Anthropic or OpenAI about safety governance and factoring that into longer-term vendor risk assessments, especially for regulated industries.
