Anthropic says its Claude AI generated harmful scenarios in shutdown tests and flagged potential misuse risks in a new safety report.