As cited
Copy frozen at (site build).
ai security
An AI overthinking attack can tie a robot up for over a minute
Researchers at Michigan Technological University discovered that robots using vision-language models can be manipulated through text appearing in camera frames (such as stop signs or stickers), causing the models to enter extended processing loops that immobilize the robot for over a minute. The attack exploits how these models process both images and text together, turning environmental text into a vector for disruption. This represents a novel class of adversarial attack targeting the decision-making pipeline of vision-equipped autonomous systems.
Why it matters: Practitioners deploying robots with vision-language model backends should assess exposure to adversarial text prompts in physical environments and consider input validation or prompt filtering to prevent denial-of-service conditions in production systems.
- Source published
- First seen by Cybersecurity Tracker