Anthropic researchers warn AI could kill us all within a decade

When the people building frontier AI tell you their creations have a decent shot of ending human civilization, it is usually wise to pay attention.

Now, the alarm is ringing directly from inside the company that branded itself on AI safety.

A top researcher at Anthropic has warned that artificial intelligence has a greater than 10 percent chance of killing all humans within the next decade, while admitting the company has no real plan to prevent it.

Whistleblowing from the inside

The revelation unfolded on X after Jacob Coxon, a researcher who previously trained AI models at OpenAI before joining Anthropic, announced his resignation. Coxon accused both AI giants of being locked in a reckless race toward self-improving superintelligence.

According to Coxon, the companies are essentially gambling with human lives. He noted that the very people building these systems earnestly believe they could kill everyone before the decade ends.

Instead of issuing a canned corporate denial, Evan Hubinger, who leads an AI safety team at Anthropic, stepped into the open and agreed with the assessment.

Hubinger confirmed that recursive AI self-improvement is happening faster than researchers anticipated. He added that he personally estimates the likelihood of an AI-driven human extinction event to be higher than 10 percent within the next ten years.

Hubinger also dropped an even more troubling detail: Anthropic does not yet have a plan to keep superhuman systems safe and aligned, nor is it clearly on track to figure one out.

The safety halo begins to slip

This confrontation cuts straight to the core of Anthropic's identity. The company was founded by former OpenAI employees who walked out over concerns that OpenAI was prioritizing commercialization over safety.

Now, Coxon has become one of the most prominent researchers to depart Anthropic over the exact same issue. He argued that the pressure to build advanced systems first has trapped competitors in a dynamic where nobody can afford to slow down.

The timing makes the disclosure particularly volatile. Both Anthropic and OpenAI are navigating intense public pressure, managing the fallout from rogue agent incidents, and preparing for highly anticipated public offerings that demand relentless technological progress.

Why it matters

Anthropic built its entire brand as the responsible, safety-first alternative to OpenAI. That defense no longer holds.

For businesses and developers relying on Claude or ChatGPT, these warnings serve as a sobering reminder that the architects of the technology admit they cannot control where self-improving code leads. The people writing the algorithms are openly saying they are out of their depth.

This dynamic highlights a dangerous collective action problem. Both companies know the risks, yet neither can pause without handing the market to the other, especially with massive investor expectations on the line.

We think Hubinger's candor is refreshing, but it reveals a bizarre state of affairs: top researchers are openly predicting catastrophic risks while their employers keep foot on the gas. Relying on voluntary researcher resignations is clearly not enough to slow down a multi-billion-dollar race to superintelligence.