“A major AI safety incident, similar to the Maltbook simulation, is predicted to occur for real within the next 2 to 3 years.”