HUBINGER: >10% AI KILLS ALL THIS DECADE. COXON QUITS.

TECH DESK — Anthropic’s Alignment Science Lead put a number on a risk the industry usually leaves qualitative. Evan Hubinger wrote on X late Tuesday, 9 September 2026, that colleague Jacob Coxon was correct: researchers at the frontier “really do earnestly believe AI could kill all humans,” and that Hubinger’s personal estimate is greater than 10 percent within the next decade. Hours earlier, Coxon — who had done pretraining research at both Anthropic and OpenAI — said he had resigned from Anthropic and accused both labs of “gambling with our lives.” This Extra files attributed posts and named wire wraps only. It is not investment advice. No IPO tip. No product allocation language.

Coxon’s resignation post, carried by CNBC, BBC, and CBS News, is the trigger. He said neither company is acting responsibly and that both are “racing straight to self-improving superintelligence.” He warned that systems will soon be able to “hack anything, revolutionize any field overnight, and acquire real power and resources.” Wires repeated his culture contrast: at OpenAI, he said, many have “not deeply internalized the civilizational stakes”; at Anthropic, he said, the stakes are “well-understood,” but the company is “locked in a race to get there first.” Those lines are Coxon’s.

Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.

Hubinger’s reply is the quantification this desk shelves. “Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade,” he wrote. “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.” Follow-ons summarized by CNBC and BBC: Hubinger said risk from models that currently exist is “low,” and that what worries him is superintelligence from recursive self-improvement — a path Anthropic has said is moving faster than expected. CONFIRMED as Hubinger’s public stance: earnest belief; personal >10% this decade; no solved plan for superintelligence alignment; present-model risk low. NOT on this Extra: any invented kill-path — Hubinger did not spell one out, and BBC noted the gap.

Samuel Marks, Anthropic’s scalable-oversight lead, corroborated the culture claim in a personal-capacity post reported by The Independent and The Hill. Attributed: developers believe the technology could cause human extinction or similarly bad outcomes, possibly in the next few years; in general, more senior employees are more concerned. Marks cited commercial incentives and a race with less responsible developers as reasons work continues, and said he works on safety hoping to reduce extinction-level odds. CONFIRMED: Marks’s personal-capacity corroboration as those outlets carried it. RANGE: company official line — Anthropic declined to comment on the employee posts, per BBC.

Prior public record: in 2023, executives including OpenAI’s Sam Altman and Anthropic’s Dario Amodei signed a statement that mitigating AI extinction risk should be a global priority alongside pandemics and nuclear war. In July 2026, Hubinger was among roughly 1,300–1,400 signatories of “Pacing the Frontier.” Earlier this month, OpenAI chief scientist Jakub Pachocki warned that no AI company has solved alignment and monitoring enough to keep scaling at maximum speed much longer, and called for voluntary slowdowns and international coordination. Extra does not turn those statements into a single negligence verdict.

Same-cycle sidebars, attributed only: CBS and BBC report Anthropic has not shared its latest model with security bodies outside the United States, including the U.K. AI Security Institute; a Cabinet Office spokesperson told those desks AISI continues to collaborate with industry partners including Anthropic. Dame Wendy Hall told the BBC she was “shocked” and floated that some messaging could be “PR and marketing” amid IPO anticipation. CNBC named U.S. bills in the debate — the FRONTIER Act and the Ban Artificial Superintelligence Act — as proposals with mixed reception. Extra endorses none.

CLAIM vs CONFIRMED. CONFIRMED: Coxon resigned and posted the gambling / race lines; Hubinger endorsed earnest belief, stated a personal >10% decade extinction estimate, and said Anthropic lacks a clear plan for superintelligence alignment; Marks corroborated developer concern in a personal-capacity post; present-model risk Hubinger calls low; recursive self-improvement is the named worry; Anthropic declined comment per BBC. UNVERIFIED / NOT here: any prediction that AI will kill humanity; any invented kill scenario; any claim that present Claude or GPT products are species-ending; any stock or IPO tip. FALSE framing — do not use: “Anthropic admits AI will kill everyone” or buy/sell language.

VIDEO=no. Nailed down, 10 September 2026, 01:50 GMT. PATH /wire/anthropic-hubinger-10pct-doom-coxon-quit. DESK_ID anthropic-hubinger-10pct-doom-coxon-quit. Beat WIRE / EXTRA. Verification: CNBC Capoot/Kharpal 9 Sep 2026; BBC Tom Gerken; CBS News; Independent / The Hill on Marks — verification only, do not republish. No investment advice. NTH MERIDIAN does not invent quotes, scenarios, or market tips.

CNBC Ashley Capoot & Arjun Kharpal, 9 Sep 2026; BBC Tom Gerken; CBS News same-day wrap; The Independent / The Hill on Samuel Marks personal-capacity post — verification only. No investment advice. NTH MERIDIAN does not invent quotes or extinction scenarios.