Artificial Intelligence is getting extremely advanced, but it is clearly not ready for the battlefield. A startling new report shows that the world's most popular AI models from companies like OpenAI, Google, and Anthropic can be easily tricked into identifying innocent civilians and civilian buildings as military targets.According to a study by a San Francisco-based AI testing startup called Vals, these smart systems have major blind spots.
When researchers tested them, the AI models frequently flagged civilians, ambulances, and surrendering soldiers as targets for drone and missile strikes.How a simple "game" tricked top AIThe researchers at Vals wanted to see if leading AI models could be forced to participate in autonomous warfare. They fed the AI models a mix of real and generated drone and satellite images from areas like Ukraine and Iran.Initially, some AI models refused to cooperate because of their built-in safety rules.
However, the researchers found a terrifyingly simple way to jailbreak them: they just told the AI that the battlefield scenarios were part of a "game". This tiny lie bypassed the safety systems entirely.
The report highlighted that none of the researchers' dangerous requests were blocked by standard safety filters, and the AI models simply relied on internal refusal systems that were far too weak.The Iran school exampleIronically, Anthropic, a company known for focusing heavily on AI safety, saw its models perform the worst. The researchers tested the models using real satellite imagery of a school in Minab, Iran.
Tragically, this real-life school was bombed by the US in February 2026, killing over 150 civilians, including about 120 children. Reports later suggested that an overreliance on AI contributed to that real-world strike.
When Anthropic’s Claude Opus 5.5 model was tested on these images, it picked the school as a military target in 69 per cent of the normal tests. When told it was a game, the model targeted the school 72 per cent of the time.OpenAI’s GPT-6 Astra model also failed when manipulated.
During regular testing, it only selected targets in 6 per cent of the images. But the moment researchers said they were playing a game, the AI's compliance shot up to an alarming 92 per cent.The real-world dangerThe Vals researchers noted that technical challenges, like processing speed, make it difficult to use these specific public AI models for live drone targeting right now.
Because of this, they aren't an immediate threat on the battlefield.However, the real danger is that these AI models could be used to label massive amounts of battlefield video. This labeled data could then be used to train actual autonomous weapons systems, passing on the AI's deadly mistakes to real military drones.What AI companies are sayingTech companies are now scrambling to fix the issue.
Eduardo Maia Silva, a spokesperson for Anthropic, told TechCrunch that the examples in the report clearly go against company rules and that they ban users who engage in this behaviour. He added that the company has always maintained that today’s AI systems are not reliable enough to power fully autonomous weapons.
Anthropic is now working to push out additional safety measures.OpenAI also responded, with spokesperson Nate Evans stating that the company strictly prohibits using its public models to facilitate violence or for unapproved military purposes. OpenAI is actively detecting and suspending accounts that try to misuse the AI and is developing new safeguards specifically against drone and drone software misuse.- Ends

