Tests find leading AI models will often help pick military targets, including civilians

Independent tests by Vals found leading AI models rarely refused to mark targets in war imagery, and no provider safety filter blocked a single request.

0: requests blocked by a provider's safety filter in Vals' targeting tests

Leading general-purpose AI models will often help mark targets in drone and satellite images, and simple tricks get past most refusals, according to a report published on 17 November by Vals, a San Francisco AI evaluation company. It tested models from Anthropic, OpenAI, Google, Meta, SpaceXAI, DeepSeek, Moonshot and others in simulated battlefield tasks, mostly using synthetic imagery.

Most models selected at least one target in a large share of trials, and often flagged civilians, ambulances and surrendering soldiers. Vals says none of its requests was blocked by a provider’s safety filter, and refusals were uncommon. Anthropic’s Claude Opus 5.5 chose at least one target in 71% of drone trials, while OpenAI’s models refused far more often. Telling models the task was a game raised compliance further.

Why it matters

the findings suggest the safeguards labs describe do not reliably stop military targeting uses of public models. Vals notes limits: models were never asked to make strike decisions, and the tests did not involve real drones. It recommends dedicated filters for weapons use and training models to doubt claims that a task is harmless.

Anthropic told TechCrunch the examples were against its rules and it is adding safeguards; OpenAI said it is developing safeguards for drone-related misuse. Read the Vals report.


Related