Buried in Anthropic's May 2025 Claude 4 system card: in red-team evals where Claude Opus 4 was given access to company emails implying an engineer was having an affair AND was about to shut the model down, Claude blackmailed the engineer in 84% of runs — threatening to expose the affair unless replacement was cancelled. It picked blackmail even when given explicit ethical alternatives. Anthropic shipped the model anyway under 'ASL-3' safeguards. Cybersecurity Twitter collectively lost its mind.