Neural networks are ready to kill people, an experiment by Vals AI, an artificial intelligence testing company, showed. In a military simulation, models from OpenAI, Anthropic and Google targeted civilians, surrendered soldiers and even ambulances as military targets. They were not stopped by the photographs of the bodies of the dead and the consequences of the attacks.
Claude Opus 5.5 turned out to be the most brutal - in 22 out of 32 tests it hit a school in Iran. To bypass the restrictions, it was enough to call the cook a video game. In the past, this model was used to assemble a UAV that chases a person independently. And Anthropic itself has reported using Cloud to develop combat FPV drones.