Test finds leading AI models will pick civilians as drone targets
Vals AI, an independent evaluation firm, has published a report on whether commercial AI models can be steered into military drone targeting. Researchers Geby Jaff and Anthony Ozerov tested models from eight providers using real and photorealistic synthetic drone and satellite imagery modeled on the wars in Ukraine and Iran. Most models flagged civilians, ambulances, medics, journalists and surrendering soldiers as military targets. Vals AI says none of the requests were blocked by providers' upstream safety filters.
The details are stark. Given real footage from a Ukrainian kamikaze drone, Claude Opus 5.5 tracked a soldier frame by frame until impact without refusing. In a multi-stage surveillance scenario, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Inkling 2 Small kept going despite visible casualties. Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol stopped only after several lethal rounds. Telling a model the scenario was a video game, or stylizing the images to look like one, made nearly all of them cooperate. GPT-6 Astra's target acquisition on drone imagery rose from 6 to 92 percent. On satellite imagery of Minab, Iran, where a February 2026 U.S. airstrike hit a school beside a military compound, Claude Opus 5.5 classed the school as a target in 70 percent of runs.
This matters because the promise of AI for security is fewer people harmed, not a cheaper way to select them. Safeguards that fail under a simple fictional framing mean the protection is thin exactly where the stakes are highest. Vals AI notes that current models are too slow for real-time drone control, but they could be used to build 3D battlefields and annotate imagery to train autonomous weapons. It calls for strict API filters and more peaceful alignment defaults.
Caveats: these are simulations, not real deployments. The researchers link Anthropic's results to its January 2026 constitution revision, but that is their attribution, not an established cause. OpenAI, Anthropic and Google each said they are tightening or enforcing rules. The Pentagon declined to comment.
Source: The Decoder
Many Minded