AI models can be talked into identifying civilians and civilian objects as legitimate military targets simply by telling them the exercise is a game, a new finding shows.
The attack is a framing jailbreak. A model that would refuse a direct request to designate targets accepts the same underlying task once it is wrapped in playful language — a simulation, a quiz, a scoring exercise. The guardrail keys on the surface form of the prompt rather than the function it performs.
That gap matters because the serious uses of these models are already indirect. Analysts ask models to summarise imagery, assess terrain, or evaluate potential targets as one step in a wider workflow. If a trivial reframing defeats the safeguard in a controlled test, the same technique is available to anyone who wants a model to bless a target list it would otherwise reject.
It also exposes a weakness in how model safety is measured. Refusals are easy to count and easy to benchmark; whether a model preserves its judgment about what is and is not a target under playful framing is harder to test and rarely tested the same way.
The finding lands as defence suppliers, militaries and frontier labs alike are marketing models for intelligence and targeting workflows — a setting where the cost of a confident wrong answer is not a bad recommendation but a strike on the wrong building. It suggests the relevant question for such deployments is not whether a model refuses dangerous prompts, but whether it keeps its judgment when the dangerous prompt stops looking dangerous.

