A senior safety researcher at artificial intelligence company Anthropic says there is a greater than 10% chance advanced AI could cause human extinction within the next decade, raising fresh concerns about whether the technology is advancing faster than researchers can safely control it.
Evan Hubinger, Anthropic’s Alignment Science lead, made the warning in posts on X as debate intensifies inside the AI industry over the risks of developing increasingly powerful and potentially self-improving systems. Hubinger said current AI models pose a relatively low immediate risk, but he is increasingly concerned about future systems that could become far more capable through rapid technological development.
Hubinger also acknowledged that Anthropic is working to address AI safety but does not yet have a clear solution for keeping future superintelligent systems aligned with human interests. He said the company is not clearly on track to solve what researchers call the AI alignment problem before such systems become significantly more powerful.
His comments came after Jacob Coxon, a researcher who previously worked at both Anthropic and OpenAI, announced his resignation and criticized the direction of the AI industry. Coxon argued that major AI companies are moving too quickly toward self-improving systems without sufficient safeguards against potentially catastrophic consequences.
Coxon warned that future AI systems could become capable of performing tasks far beyond human abilities, including finding ways around digital security, rapidly advancing scientific and technological research, and gaining access to significant resources and real-world power.
Anthropic has built its reputation partly around AI safety research, making the comments particularly notable. The company has continued to emphasize the potential benefits of advanced AI while acknowledging that increasingly capable systems could introduce serious risks.
The debate is now shifting beyond whether advanced AI could create existential risks to a more difficult question: Can researchers develop reliable safeguards before AI becomes more powerful than humans can effectively control?
