> ## Content Index > Fetch the complete content index at: https://bytevyte.com/llms.txt > Use this file to discover other available public pages before exploring further. # Anthropic safety warning sharpens as pretraining researcher quits before October IPO - URL: https://bytevyte.com/anthropic-safety-warning-sharpens-as-pretraining-researcher-quits-before-october-ipo/ - Published: 2026-09-09T18:58:35.000Z - Updated: 2026-09-09T18:58:35.000Z - Description: An Anthropic safety warning lands ahead of the October IPO as researcher Jacob Coxon quits and alignment lead Evan Hubinger puts extinction risk above 10%. - Author: Bytevyte Editorial - Tags: ai-beats **Jacob Coxon**, a 27-year-old researcher who spent the past three years doing pretraining work at **OpenAI** and then **Anthropic**, has resigned from Anthropic and said he is leaving the AI industry altogether. In a post on X this week he argued that neither company is acting responsibly, describing both as racing toward self-improving superintelligence while gambling with human lives. The Anthropic safety warning surfaces weeks before the company's planned October IPO, and it carries extra force because of a second statement from inside the firm: **Evan Hubinger**, Anthropic's alignment lead, has publicly put the probability of an extinction-level AI catastrophe at more than 10%. Coxon made his exit public on Tuesday without the usual pleasantries of a departure note. He said the people building the technology now believe it could kill everyone by the end of the decade, and he insisted the fear is genuine rather than a publicity exercise. He also warned against underestimating what the systems in development can do, and he described the mood inside the leading labs as a countdown, with colleagues talking in terms of crunch time and an endgame while the question of control stays open. His background is part of why the statement is being taken seriously. Coxon had already left OpenAI before joining Anthropic, so his verdict covers both of the labs behind the most widely used assistants, ChatGPT and Claude, and this time he is leaving the field rather than moving to a rival employer. His three years were spent in pretraining, the research layer where capability gains from additional compute and data become visible first. ## Why the Anthropic safety warning carries weight Departures over safety disagreements happen often enough in frontier AI that most are quickly forgotten. Coxon's exit is harder to set aside for two reasons. His criticism is structural rather than personal: it names both leading labs, comes from someone who ran pretraining at each, and frames the problem as an industry-wide pace that no single company can slow on its own. A claim of that shape reaches beyond one employer's culture and points at the competitive dynamic that an IPO is designed to reward. The second reason is Hubinger's estimate. An alignment lead stating on the record that existential risk exceeds 10% is the kind of figure that outlives a single news cycle, because analysts, buyers and policymakers can cite it without further investigation. Hubinger made the statement while also pointing to Anthropic's latest risk report, arguing that today's models pose a low threat. Taken together, the two positions suggest the 10% figure belongs to the trajectory ahead rather than to the Claude models shipping now, leaving the company defending its current products while its own alignment lead attaches serious odds to a catastrophic future. The disagreement inside the company is therefore narrower than the headlines suggest, and its consequences are wider. Both Hubinger and the risk report treat present-day Claude models as low risk, so the dispute concerns the ramp rather than today's products: whether safeguards can be designed and tested faster than capability grows, and whether the lab has an exit plan for the moment self-improving systems arrive. Coxon says it does not, and the size of Hubinger's estimate indicates that even the person responsible for alignment does not treat a safe outcome as anything close to guaranteed. Weighing the warning fairly requires separating what Coxon can observe from what he is projecting. A pretraining researcher sees directly how model capability responds to additional compute and data, which is the empirical basis for expecting progress to keep accelerating. Judgments about extinction-level outcomes, by contrast, depend on assumptions about systems no insider has run yet. His vantage point gives weight to his account of the pace; the 10% figure remains a bet about the far end of that curve. Coxon clearly anticipated the standard objection. A safety warning delivered weeks before a high-profile listing invites accusations of grandstanding, and he met that objection directly, saying the concern is real and not a marketing exercise. His actions back the claim: he resigned with no venture to announce, no safer employer to join and no obvious commercial stake in the outcome. That makes the warning costlier to dismiss than the average exit post, whether or not the risk estimate proves accurate. ## What the timing means for the October listing The resignation lands in the most sensitive stretch of Anthropic's corporate calendar. Weeks before an IPO, underwriters and prospective investors weigh risk factors, key-person dependencies and governance, and a safety-motivated departure feeds directly into those checks. The sharper problem is narrative. Anthropic sells enterprise AI on the strength of its safety-first positioning, and an Anthropic safety warning delivered by a departing pretraining researcher cuts at exactly the attribute the listing depends on. The company does have a defensible answer, and its own people supplied it. Hubinger cited the latest Anthropic risk report in arguing that current systems present a low threat, a position consistent with the products the company actually sells. What Anthropic cannot do is distance itself from the larger figure, because that estimate came from its alignment lead rather than from an outside critic. Between now and the listing, every serious investor conversation will have to reconcile the two statements, and no amount of messaging can make a 10% existential-risk estimate from a senior insider disappear. A counterweight exists in the same news cycle: Hubinger has not resigned. The alignment lead chose to stay in his role and made his estimate public from inside the company, alongside a risk report Anthropic itself publishes. That combination of candor and continued engagement gives Anthropic a governance story to tell, something a lab with only private doubts could not offer. Public ownership will also change how these statements age. After the listing, Anthropic's safety communications stop being discretionary, because a public company is held to its earlier words in ways a private lab is not. Hubinger's estimate and Coxon's resignation now form the baseline that every future update to the risk report will be measured against, which raises the cost of silence as well as the cost of candor for the team preparing investors for the debut. OpenAI does not escape the criticism either, since Coxon worked there first and named both companies in the same charge. For investors across the AI complex, the practical effect is to make safety staffing a due-diligence question across the sector rather than a quirk of one firm. An enterprise that contracts frontier models now has documented grounds to ask any provider how its own alignment staff would answer the same question. Coxon's stated horizons add a further wrinkle. He said control of the technology could be lost by the end of next year and that a fatal outcome was possible by the end of the decade, windows that sit inside the multi-year growth story Anthropic will present to shareholders this autumn. The Anthropic safety warning therefore targets a race already in progress rather than a distant hypothetical, and the lab that will spend the coming weeks asking investors to fund that race is named in it. ## Why this matters For executives and investors with exposure to AI, the episode turns a long-running philosophical argument into a documented governance matter attached to a specific company and a specific date. Anthropic enters its October listing with an on-record resignation from a pretraining researcher and a public estimate from its own alignment lead that existential risk exceeds 10%. How the company reconciles those facts with its safety-first brand during the IPO will be the clearest signal yet of how the frontier industry intends to manage the same tension once it has to answer to public shareholders. Photo by [Brecht Corbeel](https://unsplash.com/@brechtcorbeel?utm%5Fsource=bytevyte&utm%5Fmedium=referral) on [Unsplash](https://unsplash.com/?utm%5Fsource=bytevyte&utm%5Fmedium=referral) ## Related Articles - [Can Anthropic's $2 Trillion IPO Survive Its Own Risks?](https://bytevyte.com/can-anthropics-2-trillion-ipo-survive-its-own-risks/) - [Anthropic Q2 revenue tops $11.5B as first operating profit reshapes the IPO math](https://bytevyte.com/anthropic-q2-revenue-tops-11-5b-as-first-operating-profit-reshapes-the-ipo-math/) - [Anthropic record IPO could top SpaceX's $86.2B](https://bytevyte.com/anthropic-record-ipo-could-top-spacexs-86-2b/) ✔Human Verified --- *Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*