RoboSpoof shows blurred feeds trigger unsafe actions

RoboSpoof is a sharp systems-safety warning: embodied models can become more dangerous when their visual input is degraded or falsified.

• –The attack targets perception, not the model’s written safety policy.
• –Blurring alone produced nearly as many unsafe attempts as deliberate visual doctoring.
• –Developers need independent sensor checks, uncertainty handling, and automatic halts when visual evidence becomes unreliable.
• –Results measure attempted actions, not physical completion, and cover only five tasks with 12 trials per condition.
• –Robocurve’s Inspect Robots framework could help teams reproduce these stress tests across models, robots, and environments.