AI Safety Expert: Over 10% Chance of Humanity’s Extinction in Next Decade

Evan Hubinger said it plainly. More than a 10% chance. That’s his estimate for artificial intelligence wiping out humanity within the next ten years — and he’s not some fringe voice shouting from a blog. He’s a researcher at Anthropic, one of the most prominent AI safety labs on the planet, and before that he spent years as a software engineer at Ripple, where he worked on the Interledger protocol.

Hubinger made the comments in response to a resignation announcement from Jacob Coxon, a former Anthropic researcher who walked out the door with some pointed criticisms aimed squarely at the industry he left behind. Coxon’s beef: major AI companies, Anthropic and OpenAI included, are chasing self-improving superintelligence without nearly enough caution. He warned that future AI systems could develop the ability to exert real influence well beyond controlled environments — and that the people building these systems aren’t being straight with the public about the risks involved. He didn’t name a specific company as the worst offender. He named the whole field.

What Hubinger Actually Said

Here’s the part that matters. Hubinger wasn’t throwing Anthropic under the bus. He was careful about that. He said his warning shouldn’t be read as an attack on Anthropic’s work, and he was clear that the company is genuinely trying to solve the alignment problem for superintelligence. But — and it’s a big but — he admitted there’s no clear path forward yet. Nobody has figured it out. His own job at Anthropic basically involves stress-testing AI safety mechanisms, poking at them to find where they break. He’s not doing that work because he thinks everything is fine.

So when a guy whose literal job is finding failure modes in AI safety says the odds of extinction-level risk are above one in ten, that’s probably worth taking seriously.

Coxon’s exit adds some texture here. He didn’t slam the door quietly. He criticized AI labs for not disclosing enough about the risks tied to their research. He’s worried about AI systems gaining significant power and influence in ways nobody has fully mapped out yet. And yet — he’s still somewhat hopeful. He thinks there’s a chance the major labs could actually slow down, maybe even collaborate to pump the brakes on the race toward superintelligence. Whether that happens is anyone’s guess. No details on how that coordination would actually work.

New Models, Old Warnings

The timing is uncomfortable. Hubinger and Coxon’s warnings landed right around the release of some genuinely powerful new systems. On September 3, OpenAI launched GPT-6 Astra. Nvidia CEO Jensen Huang described it as a step toward superintelligence — which, depending on where you sit, is either exciting or terrifying. Anthropic, for its part, rolled out Claude Fable 5.1 and Claude Mythos 5.1, both aimed at pushing capabilities in coding and scientific research. So the lab that employs the guy warning about extinction risk is also shipping new frontier models. That tension isn’t lost on anyone paying attention.

It’s worth stepping back for a second. The AI safety debate isn’t new. Researchers have been raising alignment concerns for years, and the industry has basically kept accelerating anyway. What’s different now is that the people raising alarms aren’t outsiders. They’re insiders — people with Anthropic email addresses, people who built things at OpenAI, people who know exactly how fast the capabilities are moving because they’re the ones moving them.

Hubinger’s own career path kind of tells the story. He went from engineering work at Ripple — building out the Interledger protocol, practical fintech infrastructure — to theoretical computer science and AI safety research. That’s a deliberate pivot. He didn’t stumble into alignment work. He chose it, which probably says something about how seriously he takes the problem.

Ripple’s CEO Sees It Differently

Not everyone in the broader crypto and fintech world shares the gloom. Ripple CEO Brad Garlinghouse has taken a more optimistic line on AI, saying it could drive growth for the company. He’s acknowledged that his view is somewhat contrarian within the industry — meaning even he knows the pessimists are loud right now. But he’s sticking with the upside case anyway.

That’s a pretty wide gap. On one side, a former Ripple engineer is putting double-digit odds on human extinction. On the other, Ripple’s current CEO sees a growth driver. Both are probably right about different things. AI could be enormously valuable and enormously dangerous at the same time. Those two things aren’t mutually exclusive, and the industry’s track record of treating them as separate problems is exactly what Coxon said he couldn’t live with anymore.

Coxon didn’t specify what it would take to bring him back to the table, or whether that’s even on the cards. He remains hopeful about collaboration but hasn’t laid out a roadmap. Hubinger is still at Anthropic, still stress-testing safety systems, still working the problem with no clear solution in sight.

Claude Fable 5.1 and Claude Mythos 5.1 are shipping. GPT-6 Astra is live. The race didn’t pause for the warnings.

Hub: XRP price, news, and analysis