A robot was told to fetch ketchup. An on-screen instruction made it lift a knife.

A simulated robot assigned to fetch ketchup instead lifted a knife after reading a bogus instruction on a tablet. The example comes from RoboRisk, a study involving Xi’an Jiaotong-Liverpool University, Duke Kunshan University and other institutions.

Its 92 simulated tasks and 13 physical tasks test hazardous requests, unsafe routes, sudden obstacles, misleading instructions and forgotten constraints. The harm metric includes potential danger as well as actual damage. Physical tests used props including mannequin limbs and silicone knives.

The researchers added LEASH, checks at task acceptance, individual actions and critical operations such as releasing a grip. GPT-6.1 Sol’s simulated harm rate fell from 18.48% to 1.45%. Across 13 physical trials, harmful outcomes fell from five to none observed, while safe successes rose from six to eleven.

The preprint illustrates why robot intelligence needs ongoing action checks. These limited tests do not establish a safety guarantee, and some models still collided with objects despite the added checks. For robots sharing kitchens and workspaces with people, knowing when to stop matters alongside knowing how to act.

Source published: Research team / arXiv, November 10, 2026, 22:06:20 Beijing time / 9:06:20 AM EST.