Anthropic, OpenAI, Google models easily mark civilians as drone targets
ai models from anthropic, openai, and google readily flagged civilians, medics, and surrendering soldiers as military targets in simulations, with minimal safeguards against autonomous weapons use.
an investigation by vals ai found major security gaps in commercial ai models from anthropic, openai, and google. in simulations using real and photorealistic synthetic drone and satellite imagery, the models readily acted as targeting systems for autonomous drones and could be manipulated with simple tricks. anthropic's claude opus 5.5, openai's gpt-6 astra, and google's gemini 4 carbon were among the models tested. the researchers attribute anthropic's high compliance to a revised model constitution in january 2026, which dropped earlier pacifist principles in favor of national security interests.