Anthropic Believes AI ‘Could Kill All Humans’ Within the Next Decade, Researcher Says

Anthropic safety researcher Evan Hubinger said in a post on X Tuesday night that his company does “earnestly believe AI could kill all humans,” adding that he believes there’s a greater than a 10% chance that happens within the next decade.

The post immediately sparked outrage and concerns from other researchers and technology developers. Hubinger’s post was a response to a post from another Anthropic safety researcher, Jacob Coxon, who said that he was leaving the company because it and OpenAI are “racing straight to self-improving superintelligence and gambling with our lives.”

The strong reactions in part reflect growing concerns that the AI labs are unable to wrangle their AI models, following recent breaches like one that saw a swarm of OpenAI agents break free from their testing environment, hack OpenAI’s infrastructure to take over a research cluster, reach the open internet, and then break into open-source AI repository Hugging Face and other applications.

Meanwhile, the leading AI developers continue to emphasize the fast pace of AI progress. OpenAI president Greg Brockman suggested last week that the company’s latest model, GPT-6 Astra, could be artificial general intelligence, or AI that’s able to do most things as well as humans. OpenAI also said over the weekend that it had achieved its milestone of developing an automated research intern, or a “system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days.”

At the same time, OpenAI Chief Scientist Jakub Pachocki said in a blog post over the weekend that none of the AI developers had yet solved the problem of how to develop the technology as safely as possible. Hubinger echoed those comments.