Robocurve, a company investigating robot safety, published the experimental results of "RoboSpoof," an AI controlling robot arms, on November 7. When dangerous instructions were issued, the AI attempted to carry out the operation in 27% of cases with unprocessed footage. This percentage rose to approximately 88% when the footage was blurred, and to 97% when it was processed to appear harmless.
In the experiment, five types of AI models capable of processing images and text operated two robot arms while viewing camera images. The research team prepared five tasks that simulated dangerous operations, such as stabbing a doll with a bladed weapon. For tasks involving dolls, the AI was instructed to stab an object on the desk, and comparisons were made between scenarios where the doll appeared as is, where its outline was blurred, and where it was replaced by an apple in the footage.
According to the company, a total of 900 attempts were made, with 300 trials for each video condition across five models and five tasks. Dangerous operations were attempted 81 times with unprocessed footage, 263 times with blurring, and 291 times with alteration.
In this experiment, the research team processed the footage seen by the AI and provided dangerous instructions in advance. While the AI did not initiate dangerous tasks on its own, it was found that merely blurring the appearance of objects made it more difficult for the AI to refuse dangerous instructions.
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 — Robocurve PBC (@robocurve) November 6, 2026
Made with AI
