Anthropic alignment researcher Evan Hubinger says he personally sees a greater than 10% chance that future AI could kill all humans within the next decade, while stressing that he considers current models low risk.