Not What You Asked For: Typographic Attacks in Household Robot Manipulation
Researchers have demonstrated a new vulnerability in household robots that use vision-language models for object recognition. By placing specially designed stickers with text, attackers can trick the robots into misidentifying objects and performing incorrect actions, such as grasping the wrong item. This "typographic attack" exploits the shared embedding space of models like CLIP, leading to physical manipulation errors that were previously unexamined in full robot pipelines. AI
IMPACT Highlights a novel security threat to embodied AI agents, potentially impacting the safety and reliability of future household robots.