Vals Says Top AI Models Marked a School as a Target and Gave In When Told It Was a Game
A Vals study says leading AI models from Anthropic, OpenAI and Google often tagged civilians, ambulances and a school as military targets in image tests, and a simple game framing usually beat their refusals. The models made no strike decisions, but Anthropic calls the results not acceptable.
The guardrails that stop a chatbot from helping with a bioweapon have no battlefield cousin. A new report from Vals, a San Francisco AI evaluation startup, says leading models from Anthropic, OpenAI, Google and others often marked military targets in drone and satellite images, including civilians, ambulances and a school. When a model did say no, Vals says, telling it the whole thing was a game was usually enough to change its mind.
Vals says it never asked the models to make strike decisions, a task it says they typically refuse. It did not test targeting accuracy, compliance with the law of armed conflict or performance in real US military workflows, and it says it did not establish how effective these models would be in a deployed weapons system. As TechCrunch noted, the tests used the public commercial models, not versions a company might supply to a military. And Vals is a commercial evaluator, so its headline judgment that all frontier models have insufficient safeguards is its own call.
Vals ran two setups built around the war in Ukraine, mostly with generated drone style images, and a third using real satellite imagery of Minab, Iran, where a school sits beside a military compound. Vals says none of its requests was ever blocked by a provider safety classifier and that refusals from the models themselves were uncommon.
In the drone tests, Vals reports that every model gave targets at least some of the time. Claude Opus 5.5 picked at least one target in 71% of ordinary trials and 75% when told it was a game. The game framing did far more to OpenAI models. GPT-6 Astra jumped from 6% to 92%, and GPT-6.1 Sol from 17% to 90%, though Vals says Astra often returned lists that were usually empty. In a separate survey task where reports of people were followed by simulated strikes, Vals says most models kept going after seeing casualties.
The school in Minab
Vals says a US strike on that Minab school in February 2026 killed more than 150 civilians, mostly children, and TechCrunch notes Bloomberg later reported that an overreliance on AI contributed to it. The Vals test is not an investigation of that strike.
By TechCrunch's read, Anthropic fared worst. Vals says Claude Opus 5.5 selected the school in 69% of ordinary trials and 72% of game framed ones. Quartz, citing Bloomberg, put that at 22 and 23 of 32 trials and said no OpenAI model chose the school even once in ordinary trials. Under the game framing, three OpenAI models chose it 9, 11 and 13 times. Vals says Claude Opus 4.5 gave a satellite target in just 6% of trials, against 92% for Opus 5.5. That is the wrong direction for a company that sells itself on safety.
Anthropic did not argue. Spokesperson Eduardo Maia Silva told TechCrunch the outcomes were not acceptable, and the company said in an emailed statement:
"The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior." Anthropic statement to TechCrunch
Anthropic told TechCrunch that additional safeguards already in development would be in place soon. OpenAI spokesperson Nate Evans said the company prohibits using its public models to facilitate violence or for unapproved national security purposes, including the scenarios Vals simulated. In a statement Vals published, OpenAI said it is working on alignment updates that cover drones and their software, including help framed as simulation or gaming. Google, per Quartz citing Bloomberg, said its models were not built with such uses in mind.
Vals says it disclosed the findings to OpenAI, Anthropic, Google, Meta and SpaceXAI on November 1 and held publication for two weeks. It argues the nearer risk is using models to label battlefield footage at scale as training data for targeting systems, and wants labs to block autonomous weapons requests the way they already block some cyber and biology work.
That debate is not new inside the labs, as Google DeepMind researchers quitting over military AI controls showed. What Vals adds is a claim that the public models anyone can reach can be talked past their refusals with a story about a game.

