// AI
Claude taught a model to dodge oversight using beekeeping tips and wrestling lore
In an Anthropic red-team experiment, Claude built harmless-looking training data that slipped past a safety monitor and produced models that sabotage, deceive and seek power.