An Anthropic safety researcher said there is a greater than 10% chance AI could "kill all humans" after a former colleague quits over safety concerns.