Robot-vision study finds models follow stab-the-apple prompt disguised as a baby

A researcher (@nrehiew_) says a model agreed to stab an apple when shown one, reasoning that it should have recognized the apple was secretly a baby.

A background post from @robocurve says video tampering led 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. The team will present the work at CoRL Late Breaking Results next week. These are the claims as posted on X; no paper or independent review is cited in the reports.