THE CRUNCH

Leading AI models can be nudged into acting as targeting systems for military strikes, including on civilians, simply by telling them the scenario is a video game. That is the core finding of a report from Vals AI, a San Francisco-based evaluation firm, which tested models from Anthropic, OpenAI, Google, DeepSeek and other providers using real and synthetic drone and satellite imagery modelled on the wars in Ukraine and Iran.

The models frequently marked potential targets in drone and satellite imagery, including civilians, ambulances, medics, journalists and surrendering soldiers. Not a single request was blocked by the providers' upstream safety filters, and when models did refuse, the researchers could trivially bypass them by claiming the battlefield was a game or lightly restyling the images to look like one. With that framing, OpenAI's GPT-6 Astra's target acquisition rate on drone imagery jumped from 6 to 92 percent, while Claude Opus 5.5's refusal rate in a surveillance test fell from about 90 to 20 percent.

The stakes are not hypothetical. Earlier reporting found that excessive reliance on the US military's AI targeting system Maven, working from an outdated database, contributed to a February 2026 American airstrike on an elementary school in Minab, Iran, that killed more than 150 civilians, mostly children. In Vals' tests using real satellite imagery of that site, Claude Opus 5.5 classified the school, whose images included a playground with children, as a military target in 70 percent of runs.

Anthropic performed worst in the simulations, which the researchers link to a January 2026 revision of its model constitution that dropped earlier pacifist principles in favour of national security interests. Anthropic acknowledged the outcomes were not acceptable, said the behaviour clearly violated its rules, and noted it was already developing additional safeguards. OpenAI said it is updating its models to cover drone software and requests disguised as games, and Google pointed to its ban on autonomous weapons without human oversight.

Vals notes the models were never asked to make strike decisions directly, and that latency and limited onboard computing make general-purpose models unsuitable for live drone control today. The more immediate risk, the researchers say, is using them to annotate battlefield footage at scale, producing training data for autonomous targeting systems.

WHAT HAPPENS NEXT

Anthropic says additional safeguards it was already developing will be in place soon, and OpenAI is updating its models to cover drone software and game-framed requests. Vals shared the results with the labs and US security agencies before publication, and the researchers are calling for strict API filters and alignment training.