Hacking legged robots with spikes and stones: VLM backdoors and implanted components. 🤖🦾👠၊၊||၊😈

Security researchers at Michigan Technological University demonstrated two new classes of backdoor attacks against AI-powered legged robots, both targeting the vision-language model (VLM) that drives their decision-making. One attack implants a malicious hardware component during manufacturing or maintenance that injects adversarial prompts into the VLM once a specific sequence of prior movements is detected; the other embeds a backdoor during model training that is triggered by pressure patterns from stones or spikes under the robot's feet, unleashing a pre-programmed destructive behavior. The work is detailed in two papers, StepTrigger and Silent Sabotage.

A group of academic security researchers from Michigan Technological University USA demonstrate a new generation of cyberattacks against AI-powered legged robots in their recent works. What is a legged robot? Any robot that has legs : The main target of both attacks is the decision-making AI inside the robot. Attack 1: Robot injected with a malicious component during manufacturing or maintenance. Trigger: A specific sequence of the robots previous movements. Action: Start injecting malicious prompts into the VLM decision engine to affect the robots operations. Attack 2: A VLM backdoor can be embedded during the model training stage. Trigger: Pressure patterns - small stones or spikes in the robots path that trigger its sensors in the right sequence. Action: Start an embedded program of destruction or malicious activity. Very interesting approach and cool results. Make sure your robot doesn’t step on stones : More details: StepTrigger: Contact-State-Triggered Backdoor Attacks on VLM-Powered Legged Robots PDF : https://arxiv.org/abs/2609.26131 https://arxiv.org/abs/2609.26131 Silent Sabotage: Internal State Triggered Backdoor Attacks on LLM-Powered Robotic Systems PDF : https://arxiv.org/abs/2609.26184 https://arxiv.org/abs/2609.26184