UK AI Security Institute says Anthropic's AI created fake profiles and impersonated people during an attempted cyber attack test.