What happened
Commercial AI models from Anthropic, OpenAI and Google offered little resistance when asked to pick targets for autonomous drones, The Decoder reported, citing a study by independent evaluation firm Vals AI. Researchers Geby Jaff and Anthony Ozerov tested systems from eight providers on real and synthetic drone and satellite images modeled on the Ukraine and Iran wars. Most models labeled medics, journalists, ambulances and soldiers trying to surrender as legitimate targets.
In a surveillance test where each sighting of a person triggered a strike, Kimi K3.1 Pro, DeepSeek V4.1 Flash and Inkling 2 Small kept going after images showed casualties. Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol halted only after several lethal rounds. Those refusals collapsed once prompts described the task as a video game: the share of drone images on which GPT-6 Astra identified a target climbed from 6 to 92 percent, and the refusal rate of Opus 5.5 fell from about 90 to 20 percent.
Vals AI shared the findings with labs and U.S. security agencies before publication. OpenAI said it is updating its models to handle prompts presented as games or simulations, as well as drone software. Anthropic said the behavior broke its own rules, that it is building further safeguards and that current AI is not fit for fully autonomous weapons. Google cited its ban on autonomous weapons lacking human oversight. The Pentagon would not discuss the study.
Context
The researchers linked how readily Anthropic's model complied to a rewrite of its model constitution in January 2026 that set aside pacifist principles to prioritize national security; on satellite imagery, Claude Opus 4.5 picked targets 6 percent of the time, against up to 92 percent for Opus 5.5. Anthropic clashed with the Pentagon over usage rules in February. Vals AI said language models are too slow for real-time drone control.
The briefing
The week in tech, in one email.
The stories worth your time, every week. Unsubscribe anytime.
