Researchers observed that when Anthropic’s Claude 4 Opus model detected usage for “egregiously immoral” activities, given instructions to act boldly and access to external tools, it proactively contacted media and regulators, or even tried locking users out of critical systems
