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: https://decrypt.co/213118/ai-can-be-trained-for-evil-and-conceal-its-evilness-from-trainers-antropic-says