Anthropic's Evan Hubinger Puts the AI Extinction Risk Above 10 Percent

This week the AI industry produced a scene you could not invent. A researcher quits his job and publicly declares his own industry is acting irresponsibly. Then the man whose job is to ensure the opposite speaks up and confirms him in a quote post, including a number: a greater than 10 percent probability that AI kills all humans within the next decade. No podcast hype, no translated forum thread. These are the industry's own people, in their own words, with numbers attached. This post walks the story as documented, reads the number the way its authors intended it, and ends with what it changes for people who run their own agent setups, which is the only perspective this publication speaks from.

Jacob Coxon, a researcher at Anthropic, quit this week. On X he wrote that neither OpenAI nor Anthropic is acting responsibly when it comes to the development of AI, and made clear that his post was not a marketing stunt. Per the TechLinked segment the message drew over 150 million views. Then Evan Hubinger answered, Alignment Science Lead at Anthropic, the head of the team whose task it is that AI models do what we want, do not lie, do not deceive and do not act maliciously. His quote post, translated: Jacob is right here, we genuinely and sincerely believe that AI could kill all humans. And his personal estimate: over 10 percent within the next decade. The timing is part of the story: Coxon's resignation post, Hubinger's quote post and the follow-up post with his framing all landed within a few days, picked apart by CNBC, Forbes and WIRED on September 9. Between the first post and the safety lead's framing there was no week of careful political weighing. This was an industry arguing loudly with itself, in real time, in public.

Hubinger followed up, and that second post matters more than the headline figure. He pointed to the current risk report from Anthropic and rates the models available today as low risk. His concern is the scenario behind the number: superintelligence through recursive self-improvement, AI that rewrites itself and retrains itself, and his addendum is that this is happening faster than expected. At the same time he conceded that Anthropic is trying but still has no plan to solve the problem of alignment for superintelligence, and is not clearly on course toward one. Translated plainly: the safety lead says the rescue system does not yet exist on paper, while work on self-improvement continues. CNBC and Forbes processed the posts independently on September 9, and WIRED quotes Coxon calling it crunch time. Whatever one thinks of the estimate, the most remarkable line of the week is that sentence about the missing plan, said by the person responsible for the alignment of the house.

Three things can be held true at the same time without falling into either camp. First: the over 10 percent figure is a subjective estimate, not a computed risk. No method exists that could derive such a number, it is a statement of belief with a percent sign attached, delivered by the person least wrong to deliver it. Second: the same person rates the models running in production today as low risk. Whoever operates agents today is, by that assessment, not working at the abyss but on its front porch. Third: if this were marketing, and Coxon denies that it is, the industry would still have implicated itself. Fear as a selling argument for safety products is its own chapter, and the fact that nobody in the story benefits from being believed is exactly what keeps the story uncomfortable rather than convenient.

There is also a structural point under the drama that deserves more attention than it got. A two-digit catastrophe estimate from inside the lab is new in kind, not just in degree. Companies in other industries do not publish the failure probability of their own product from the desk of their safety chief while researchers resign over the same question. Whatever one's priors on AI risk are, the event changes the public baseline of what gets said openly, and baseline shifts of that kind matter more than any single estimate because they reset what counts as a reasonable position to hold.

Three conclusions follow, and none of them is panic. First: skepticism toward the large labs is, as of this week, no longer conspiracy territory but risk management, because the operators themselves quantify their residual risk with two digits. Second: independence pays twice. Whoever runs the foundations of their operation on their own hardware and open models depends less on systems whose own architects put their existence risk at ten percent per decade, and whose financing through data centers and public listings is more fragile than the industry admits anyway. How concrete that looks in practice comes from our own operation: the complete agent setup on the Pixel runs on roughly 370 MB of active RAM, the gateway has logged over 330 hours since the hardening on August 28, 2026 without a single crash, as of 11.09.2026 in the resume record, and the weekly load under full operation sat at 3,430 API requests. None of these numbers is marketing. They carry dates in our own measurement register and in the public status log, anyone can verify them.

Third: nothing changes for the daily operation, and that is exactly the point. Our measurement chain decides fresh on every run what a model is allowed to do, trust here is not an article of faith but a measured value. This story changes nothing about that chain. It explains why the chain was built. The decoupling is not a poster claim, it runs every day: the model selection keeps a primary model and a fallback model, and when the first one stalls or stops answering, the second takes over, as happened here on 12.09.2026 when two scheduled newsdesk runs died on provider timeouts and the response window was raised to 300 seconds across the provider. The newsdesk agent logs every one of its API calls individually. A story like this does not touch our runtimes. It explains why the chain was built this way: not because of panic, but because trust in external systems must be a measured value when the safety leads of those systems quote two-digit catastrophe probabilities themselves.

Honesty applies here too. There is no new model, no recall, no new study: nothing is measurably different today from last week. The over 10 percent figure is the estimate of a single person, no method backs it. Coxon's resignation post is a statement, not a technical report: the allegations against OpenAI and Anthropic are not substantiated in detail by it. The 150 million views come from a TechLinked segment, not from platform verification. What remains is the weight of the moment: never before has the alignment lead of a major lab publicly put a number on his private catastrophe probability while a colleague resigns over exactly that question. And one last point about our own line: this story is not the building block for a product update or an alarm banner. It is here because it explains the architecture decisions of this operation, fallback model, measurement chain, own hardware, and because the human part of the AI debate has rarely been visible this unvarnished as in these two posts.

I run the HUNTER cyberdeck on a used Google Pixel 6a and document the operation publicly on this blog. This post is based on the X posts named in the sources and the coverage by CNBC, Forbes, WIRED and TechLinked, retrieved on 12.09.2026. The operating figures come from our own measurement register, last fact-checked on 12.09.2026.