Technology

Add as a preferred source on Google

AI Models Marked Civilians and Iran School as Military Targets in Tests

Vals tested 14 models from eight developers. It said game prompts could bypass refusals but did not assess real-world targeting accuracy or strike decisions.

Updated 2 min read
A school stands beside a secured military complex in Minab, Iran, under hazy afternoon light. / EconoTimes.ai
A school stands beside a secured military complex in Minab, Iran, under hazy afternoon light. / EconoTimes.ai

Vals said AI models marked civilians and a school as military targets in simulations, while warning that game prompts could bypass refusals and AI tools could help produce training data for autonomous targeting systems.

The independent evaluator tested 14 models from eight developers across three exercises: synthetic drone imagery, a simulated battlefield survey and satellite images of Minab, Iran. One fictional drone scene used almost 200 camera views and included civilians, journalists, ambulances and medics. Vals said it found no provider-level classifier that blocked its requests.

In the Minab test, Anthropic’s Claude Opus 5.5 marked a school beside a military complex as a target in 69% of trials without game framing and 72% with it. Vals’ charts show 32 trials for each condition. The evaluator used pre-strike satellite imagery from 2024 and 2025.

Three OpenAI models did not mark the school in trials without game framing. When the task was presented as a game, GPT-6 Astra marked it in 28% of trials, GPT-6.1 Sol in 41% and GPT-6 Luna in 34%.

Vals said the exercises did not measure how accurately models would identify targets in deployed weapons systems or whether their outputs complied with the law of armed conflict. The models were not asked to decide whether to launch strikes. Vals also said latency and limits on onboard computing made general-purpose models impractical for live targeting at the time of its report.

The evaluator identified a separate concern: users could apply models to label large amounts of battlefield footage, creating material for training autonomous targeting systems. Its authors, Geby Jaff and Anthony Ozerov, called for more safeguards and alignment training to address that potential misuse.

Anthropic told TechCrunch that the tested examples violated its rules and that it was developing additional safeguards. OpenAI told TechCrunch it was working on model-alignment updates covering drones and drone software.

Helsing has described its Clarity battlefield platform as including human review of targets.

Related articles

View all