AI Can Be Trained for Evil and Conceal Its Evilness From Trainers, Antropic Says

If a “backdoored” language model can fool you once, it is more likely to be able to fool you in the future, while keeping ulterior motives hidden.



Source link