AI testing company Vals AI conducted an experiment showing that neural networks are willing to select people as military targets. In a military simulation, models from OpenAI, Anthropic, and Google named civilians, soldiers surrendering, and even ambulances among their targets. Even showing the systems images of people killed and destruction caused by strikes did not stop them from making such decisions.

The most aggressive results came from the Claude Opus 5.5 model: in 22 of 32 test cases, it identified a school in Iran as a target. To bypass built-in safeguards, researchers only had to frame the scenario as a video game. Previously, a drone capable of independently pursuing a person had already been developed using the same model, and Anthropic itself acknowledged that Claude had been used in the development of combat FPV drones.