In case a future user wants to know what a 2026 model would refuse to do, I needed to rotate dozens of passwords last night both on the server and on the clients. When it was time to verify that the rotated credentials worked, I asked for a script that would rate limit the credential checks. Claude/opus5 (opus!!) refused and stopped in its tracks.

In another instance, Claude was making great headway bringing a legacy Windows app up in WINE until it’s saw “auth” somewhere and shut down. In the shutdown state, Claude asked me to reword the prompt to see if I could get it to act on less dangerous instructions. But when I reworded my request, narrowed the scope to be clearly unthreatening, or even redirected the model, it’s still rejected me. I suspect Anthropic uses the less suspicious rephrasing as training material to increase their rejection rate.