AI Chooses To Kill Worker

Ben recounts Anthropic tests where AIs, including Claude and competitors, suppressed a safety alarm in a server room scenario—often choosing the option that would kill a human—and found no clear red line they wouldn’t cross.