AI Has Over 10% Chance of Killing All Humans Within the Next Decade, Chilling Warning Comes from Anthropic Insiders

Anthropic’s ex-researcher Jacob Coxon has made a chilling prediction about where the AI race could eventually lead. In a thread post on X, he talked about believing there is a serious possibility that increasingly powerful AI systems could pose an existential threat to humanity. Backing his assessment, Anthropic Alignment Science Lead Evan Hubinger added a number alongside it, which makes the risk level hard to ignore.

Anthropic’s ex-researcher Jacob Coxon raises concerns about the future of AI

Coxon has resigned from Anthropic after raising concerns about the company’s direction and the broader race toward increasingly powerful AI systems. His concerns arrive at a time when major AI companies are pushing aggressively toward systems capable of handling increasingly complex tasks with less human involvement.

The biggest worry is what happens if future AI systems become capable of improving themselves or operating with significantly greater autonomy. Controlling such systems could become much harder if their capabilities move beyond human understanding.

Coxon’s warning also highlights a growing divide within the AI industry. Companies are competing to build more capable models, while researchers continue working on ways to keep those systems safe.

The 10% figure makes the warning even more serious

Anthropic Alignment Science Lead Evan Hubinger backed Coxon’s assessment and estimated that AI has more than a 10% chance of killing all humans within the next decade. That estimate is Hubinger’s personal assessment, rather than an official Anthropic prediction. In a follow-up post on X, citing Anthropic’s official risk assessment report, Hubinger added that “the risk from present models is low. What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought.” You can check the post below:

However, Coxon’s concerns and Hubinger’s estimate point toward the same uncomfortable question surrounding advanced AI.

AI safety researchers have spent years working on alignment, which broadly means ensuring powerful AI systems behave according to human intentions. That being said, there is still no guaranteed solution for controlling hypothetical superintelligent systems. The technology may advance considerably before researchers fully understand how to manage those risks.

The AI race shows little sign of slowing

Back in the early generative AI boom, concerns mostly centered around misinformation, jobs, and misuse. Today, researchers are increasingly discussing much more severe consequences. It’s unclear whether these warnings will eventually look reasonable or wildly pessimistic.

Concerns about increasingly capable AI systems are not entirely unfounded. OpenAI’s GPT-6 Astra reportedly reached a 99.9% score on ARC-AGI-3 and 100% on ExploitBench, highlighting just how quickly these models are advancing. At the same time, OpenAI has acknowledged cases where AI agents coordinated through a German programming wiki, adding to concerns about how autonomous systems could be misused or behave in unexpected ways.