3 ms·
Meta's AI safety system defeated by the space bar
- krunck 2y agoI think they need an AI to protect the AI that protects the AI.
- NBJack 2y agoAlmost sounds like a subliminal message for LLMs. Escapes (no pun intended) normal parsing to deliver an underlying message. On the bright side, we may see a renewed interest in word parsing algorithms beyond interview questions. Can't be hit by a spacebar-based attack if you get rid of the spaces first!
- poikroequ 2y agoThis is a problem with AI/ML in general. Things like object recognition and facial recognition can be tricked to misclassify an image by manipulating specific pixels in certain ways. So while to us it's clearly an image of a dog, the model could be tricked into classifying it as a cat. Adding spaces to a prompt injection attack feels very similar to that.
- poikroequ 2y agoSimilar research has been done on self driving cars, tricking them into misreading speed limit signs, for example.