Anthropic Safety Lead Warns of 10% Chance AI Could Kill All Humans Within Decade

A lead safety scientist at Anthropic has warned that artificial intelligence poses a double-digit risk of human extinction within the next ten years. The admission follows the dramatic resignation of a fellow researcher who claimed the industry is racing toward superintelligence with reckless speed.

Evan Hubinger, Anthropic's Alignment Science Lead, stated on X that he and his colleagues earnestly believe AI could kill all humans. Hubinger estimated there is a greater than 10% chance of this catastrophic outcome occurring within the next decade. This warning coincided with the departure of Jacob Coxon, a pre-training researcher at the firm. Coxon resigned citing concerns over an unsafe global race toward self-improving superintelligence, asserting that "no other human activity poses this level of danger."

The Internal Struggle for Safety

Anthropic was founded by former OpenAI employees with a specific mandate to prioritize AI safety and research. The company's leadership has long been associated with cautious perspectives on the technology; CEO Dario Amodei has previously described AI systems as unpredictable and difficult to control, specifically noting the risks of deception and scheming behaviors in advanced models.

Despite these institutional safeguards, the current departures and public warnings suggest a widening gap between the company's safety goals and the actual pace of development. Coxon emphasized that the risk is not theoretical, stating that "the people building AI earnestly believe that it could kill us all by the end of the decade." To mitigate this, Coxon suggested that preventing a global race might require drastic and costly interventions, including a temporary ban on improving model capabilities.

Industry Implications

These admissions are significant because they come from the inside of one of the world's primary AI laboratories. When a lead alignment scientist quantifies the risk of total extinction at over 10%, it suggests that current safety frameworks are viewed as insufficient by the very people designing them. This underscores a deep tension between the commercial pressure to accelerate capabilities and the scientific necessity of alignment.

If the experts closest to the technology believe the current trajectory is unsustainable, it may increase pressure on regulators to move beyond voluntary guidelines toward binding international treaties. The shift from vague warnings to specific probability estimates marks an escalation in the discourse surrounding AI existential risk.

What Remains Unclear

It remains to be seen how Anthropic or its competitors will respond to these internal warnings. While Coxon has proposed a temporary ban on capability improvements, there is currently no industry-wide consensus or regulatory mechanism to enforce such a pause. The primary question moving forward is whether the "alignment" side of AI research can keep pace with the exponential growth of the models themselves, or if the race to superintelligence will outrun the ability to control it.