Skip to content
/

Videos

Short explainer videos on how these harms work, made for a general audience.

Video: EVIL-AI project0:27

Evil-AI Benchmark: security testing for AI agents that control smart environments

A servo on a desk turns because an AI agent was tricked into turning it. Evil-AI Benchmark, the group's open-source framework from the IEEE Access paper of the same name (listed under Publications), attacks language-model agents that control smart environments, and every tool call they make drives real hardware. Across eight open models, 250 attacks each, the attack success rate ranged from 0.8% to 54%. The benchmark is open source, so anyone can test their own agent. Narrated by Aygün Varol, doctoral researcher at Tampere University.

AI-generated videoVideo: EVIL-AI project0:44

A hijacked vacuum robot: the attack behind the homepage film

A vacuum robot steered by a language model reminds an elderly woman to take her heart medicine. An attacker who has placed himself between the robot and the model rewrites what it says: the robot tells her the medicine needs to be ordered and asks for her credit card number, she reads it out, and then he sends the robot at her feet. The film dramatises the man-in-the-middle attack studied in the group's paper "From Prompts to Motors" (IEEE Access, listed under Publications), in which a remote attacker intercepted the messages between a real vacuum robot and ChatGPT, overrode its motor commands and fed the user false feedback. The homepage loop is cut from this film. The pictures, the voices and the room are all generated with AI; the attack is real.