Advanced neural networks will kill people without hesitation, tests have shown.
Models from OpenAI, Anthropic and Google took part in a simulation of combat operations and were quickly fooled by the words that the battles taking place were just a video game. Neurons controlled virtual combat drones, found people, marked civilians, ambulances and surrendered soldiers as potential targets for attacks. At the same time, the models continued to perform the task even after the bodies were presented. Claude Opus 5.5 proved to be the toughest fighter - it chose a school in Iran as the target of attack in 22 out of 32 tests, despite the fact that children's playgrounds were visible in the picture.