There is a more than 10% chance that "AI could kill all humans" in the next decade, according to a top safety researcher at Anthropic.