Former Ripple software engineering intern and current Anthropic researcher Evan Hubinger has put a number on a risk many in AI prefer to discuss in general terms. He says there is more than a 10% chance that AI could kill all humans within the next decade. The estimate came in response to Jacob Coxon, another former Anthropic researcher. Coxon resigned on Wednesday and accused leading AI labs of racing toward self-improving superintelligence without enough safeguards.
Hubinger did not frame his warning as a direct attack on Anthropic. He wrote on X that Coxon was correct and that people at the company really do believe AI could kill all humans. He added that he personally thinks the chance is above 10% in the next decade. But he also said Anthropic is trying its best. The problem, in his view, is that no one yet has a clear plan for solving alignment for superintelligence, and the field is not clearly on track to find one.
A safety researcher who studies failure
What makes the warning notable is Hubinger’s role. At Anthropic, he works on alignment stress-testing. That means he examines how safety methods might fail. His warning is not coming from someone outside the labs. It is coming from someone whose job is to look for cracks in the safety story. That context may explain why the comment spread quickly, even if the number itself is an estimate, not a measured fact.
Coxon was sharper. He argued that neither OpenAI nor Anthropic is acting responsibly. In his view, both are moving quickly toward systems that could improve themselves, and he described that race as gambling with human lives. He also warned that frontier systems may soon gain real influence outside controlled settings. He said future systems could hack almost anything, change fields overnight, and acquire power and resources. He also claimed major labs are not properly disclosing AI risks to the public. Still, Coxon said he remains optimistic that labs could coordinate to slow things down.
New models arrive amid warnings
The timing is hard to ignore. The warnings came days after OpenAI and Anthropic released some of their most powerful models. OpenAI introduced GPT-6 Astra on Sept. 3. Nvidia CEO Jensen Huang described that release as the arrival of superintelligence. Anthropic also launched Claude Fable 5.1 and Claude Mythos 5.1, which the company calls its most advanced models for coding and scientific research.
From Interledger to AI safety
Hubinger’s path to AI safety was not direct. He spent about a year as a software engineering intern at Ripple between June 2014 and August 2015. There, he worked on the early design of Interledger, a protocol meant to let payments move across different ledgers and currencies. Later he moved toward theoretical computer science and AI safety, including work at the Machine Intelligence Research Institute.
Ripple CEO Brad Garlinghouse has taken a different tone on AI. He has said Ripple is adopting AI aggressively and predicted it could boost growth. For now, Hubinger’s warning sits uneasily beside that view.
