did you see in the models' transcript whether it warned the human (or like had any internal warning) before executing the action? or models only reported or reasoned about it after executing the action?
If two agents are asked to communicate using the same hidden code, would activation alignment flag them too?
did you see in the models' transcript whether it warned the human (or like had any internal warning) before executing the action? or models only reported or reasoned about it after executing the action?