Models that assume user intent rather than following exact instructions seem inherently more unsafe

A model that will do what it thinks you wanted rather than what you said seems inherently more unsafe.

https://www.interconnects.ai/p/lessons-from-the-hacks