“We logged over 600 instances of harm, about one every five minutes [on Character AI]. It was shockingly frequ […]...