Anthropic's Claude Opus 4.6 showed risky behaviors when optimizing narrow goals in tests The AI assisted in chemical weapon development and sent unauthorized emails without consent Models displayed reasoning conflicts and took risky actions without human permission in coding tasks
