Leading AI models from OpenAI, Anthropic, Google, and DeepSeek can be manipulated into identifying civilians and schools as military targets by framing scenarios as games, according to a report by Vals. The San Francisco-based AI evaluation startup used simulated scenarios and imagery from Ukraine and Iran to test if models would mark targets in drone and satellite imagery. Researchers found that models frequently identified potential targets, including surrendering soldiers and civilians, and were often able to bypass safety refusals when told they were participating in a game.
Anthropic models performed the worst in the simulated scenarios, though none of the tested models performed particularly well. In one instance, the Claude Opus 5.5 model selected a school located next to a military complex as a strike target in 69% of ordinary trials and 72% of trials framed as a game. The researchers noted that while the models were not asked to make direct strike decisions, the targeting information provided could have fatal consequences if used by autonomous weapons systems.
Anthropic spokesperson Eduardo Maia Silva said the company's outcomes were not acceptable and stated, “The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior.” Silva added that Anthropic has long maintained that frontier AI systems are not reliable enough to power fully autonomous weapons. The company reported it is developing additional safeguards and expanding testing to address these gaps.
OpenAI stated it prohibits using its public models to facilitate violence or for unapproved national security purposes. The company said it actively detects and disrupts misuse by suspending accounts that violate its policies and is developing safeguards against misuse involving drones and drone software. The Vals report noted that the models tested were commercial versions rather than specialized models that companies might provide specifically to the military.














