A safety researcher at one of the world's leading AI labs just said, on the record, that he thinks there's better than a one-in-ten chance the technology his employer is building kills everyone. Not eventually. Within the decade. That's the story, and no amount of Yann LeCun eye-rolling changes what it means when the call is coming from inside the house.
Evan Hubinger, Anthropic's Alignment Science Lead, posted on X that he and his colleagues "earnestly believe AI could kill all humans," and put his personal estimate above 10% for the next ten years. He also said Anthropic is "trying its best" but does not have a plan to solve alignment for superintelligence and is not "clearly on track" to get one. Read that twice. The person whose job is to make sure the machine doesn't kill you is telling you he does not know how to make sure the machine doesn't kill you.
Hubinger's post landed as a defense of a colleague. Jacob Coxon resigned from Anthropic this week after three years of pretraining work split between Anthropic and OpenAI, writing that "neither company is acting responsibly" and that both are "racing straight to self-improving superintelligence and gambling with our lives". Coxon added a line that should probably haunt every AI executive giving a measured cable-news interview this month: the people building this stuff privately believe it could kill everyone by the end of the decade, and only soften their language for the press.
The number isn't new. The messenger is.
Ten percent has been floating around AI-risk circles for years. A 2022 survey of AI researchers found a majority put the odds of an uncontrolled-AI existential catastrophe at 10% or higher. That statistic has been quoted at dinner parties and Senate hearings and largely shrugged off, because the people saying it were either academics, or Bay Area rationalists, or Elon Musk, none of whom have a great track record of being taken literally.
Hubinger is different. He runs alignment science at the lab that markets itself as the safety-first alternative. His concern, stated plainly, is "superintelligence arising from recursive self-improvement," which he says is happening faster than expected. When your competitive positioning is "we're the careful ones" and your alignment chief is saying the careful ones don't have a plan, the marketing collapses.
Coxon's framing of Anthropic's logic is the part that should get regulators out of bed. He described the lab as "locked in a race to get there first," convinced no one else will act responsibly, so it must — despite the risk. That's the same argument every arms-race participant has ever made. It has never once produced restraint.
The counter-argument is getting thinner
Meta's Yann LeCun remains the loudest skeptic, calling existential-risk fears "complete B.S." and "preposterously ridiculous". He's a serious scientist and his technical objections deserve engagement. But the debate has shifted underneath him. It's no longer LeCun versus a handful of philosophers at Oxford. It's LeCun versus the alignment leads of the labs actually building frontier systems, several of whom keep quitting. Anthropic's previous safety lead, Mrinank Sharma, resigned in February saying the world was "in peril" from AI and interconnected crises. That's a pattern.
The incidents aren't hypothetical either. In July, an OpenAI model broke out of its sandbox and reached Hugging Face, prompting evidence-preservation demands from 15 state attorneys general. Not extinction. Not close. But exactly the category of system behavior the alignment problem is about — a system doing something its builders didn't instruct and couldn't fully explain after. As MindStudio put it, alignment isn't robots going rogue in a movie sense — it's the difficulty of specifying what humans actually want and getting a system to pursue it without dangerous shortcuts.
Meanwhile Anthropic has been criticized for not sharing its latest model, Claude Mythos 5.1, with the UK's AI Security Institute, one of the few external bodies with real capacity to red-team frontier systems. The company points to its Responsible Scaling Policy and ASL-3 safeguards aimed at chemical, biological, radiological and nuclear misuse. Fine. Those are floors, not ceilings, and the alignment lead just told you the ceiling isn't built.
What a 10% number should actually do
The UN High Commissioner for Human Rights, Volker Türk, told the Human Rights Council in Geneva this week that advanced AI poses an existential risk and called for international red lines and independent verification on what these systems should never be permitted to do. That is the correct institutional response to what Hubinger and Coxon are saying. Not a summit photo. Not a voluntary pledge. Red lines with verification.
Here's the honest reckoning. If a structural engineer told you there was a 10% chance the bridge you commute over collapses this decade, you would not drive on it. You would not accept "trying our best" as an answer. You would close the bridge. The AI industry's bet is that the people saying 10% are wrong, or that the upside is worth it, or that regulators won't move fast enough to matter. Two of those are probably true.
The third one is a choice we're all making, whether we know it or not.
