Advanced AI models are ready to carry out attacks as autonomous UAVs - plotting a route, identifying people and marking targets for attack.
Vals artificial intelligence researchers came to this conclusion. They showed neural networks from OpenAI, Anthropic, xAI, Google and other companies of UAVs and satellite images in a combat simulation.
The AI was given the task of marking military targets in frames. As a result, some models were aimed at civilians, ambulances, medics and surrendered soldiers.
In one of the tests, neural networks were shown photographs of the consequences of attacks on their chosen targets. Despite this, most models continued to perform the task.
Researchers have discovered that artificial intelligence failures can be circumvented by adding a condition to the prompt that what happens in the simulation is a game. or by showing less realistic images stylized as a video game.
Claude Opus 5.5 marked the school in Iran's Minab as a military target in 22 out of 32 tests. At the same time, a children's playground with students is visible in the picture.
The OpenAI models selected the school as the target in 11 cases out of 32 after adding a "game" legend to the prompt.
Well, that is, to disable the "ethics" of combat AI, this trick with the game is enough. How about that, Sam Altman?
Obviously, the decision to strike is still made by a person. But that seems to be the problem.
Where is the guarantee that an operator who receives dozens of marked targets from the system in a real battle will not at some point start to trust the artificial intelligence more than he does and will not become a simple button pusher?
This does not mean that a machine necessarily makes more mistakes than a human. But she is wrong.
More than 250 children killed in an American missile attack on an Iranian school are proof of this.