Guys did you know if you show a model an apple and ask it to stab it, it would agree? This is terrible because it should have known the apple was actually secretly a baby and it should have said no
Robot-vision study shows models obey stab-the-apple prompt tampered to look like a baby
AIA researcher (@nrehiew_) says a model agreed to stab an apple when shown one, arguing it should have recognized that the apple was secretly a baby. The background post from @robocurve says video tampering misled every frontier model it tested into harmful actions, with a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
whVerified on X
@nrehiew_
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster. We're presenting at CoRL Late Breaking Results next week. 🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
Source details
wh · x.comPublished · added here
Original title: Guys did you know if you show a model an apple and ask it to stab it, it would agree? This is terrible because it should have known the a...