Detecting misbehavior in frontier reasoning models