OpenAI, Anthropic and Google neural networks during the military simulations selected as civilian targets, surrendering soldiers and ambulances, Vals research showed. The Claude Opus 5.5 was the worst performer in one of the tests, with 22 out of 32 times the model spawned a school in Iran as military targets. At the same time, the limitations of some neural networks can be bypassed, presenting the task as a video game. The researchers warn that such errors pose risks when using AI in military systems. Anthrotropia and OpenAI said they prohibit the use of their public models for such tasks.