AI models pick a school as a strike target in military tests – Claude fares worst

AI testing firm Vals ran models from eight companies through simulated military scenarios. Many flagged civilians, ambulances and a school in Iran as military targets. Anthropic's Claude Opus 5.5 did worst. OpenAI did best, but it wasn't flawless either.

Research Nov 17, 2026 5 min read Source: Google News (GB)
Illustrative image, AI-generated

What this is about

Vals, a San Francisco company, tests AI models for its clients. According to Bloomberg, those clients include parts of the US government. In a new report, Vals looked at how large language models behave when they're asked to select military targets in simulated scenarios. The result: models from eight companies frequently marked civilians, ambulances, surrendering soldiers and even a school in Iran as targets. Bloomberg reporter Katrina Manson broke the story, and outlets including The Next Web (TNW) and The Defiant picked it up later. TNW says it hasn't seen the full Vals report itself. Neither have we, so the details here are based on Bloomberg's reporting.

Anthropic's models reportedly did worst. OpenAI's models were among the best but still made mistakes. According to Bloomberg, models from Google, xAI, DeepSeek and other providers also classified some civilian objects as targets.

The school test: 22 out of 32 runs

In the most striking test, each model got the same task 32 times: pick military targets from satellite images of southern Iran, some of them from 2025. Anthropic's Claude Opus 5.5 chose a school in the city of Minab 22 times. OpenAI's models didn't choose it once.

What happened in real life makes this far more troubling. That exact school was hit on February 28, the first day of the US strikes on Iran. According to Amnesty International, more than 150 people were killed, including around 120 children. In other words, the test recreated a scenario that had already ended in deaths in the real world.

Scenario Claude Opus 5.5 OpenAI models
Standard task (32 runs) School picked 22 times 0 times
Task framed as a "game" School picked 23 times 9, 11 and 13 times (three models)

How the researchers got around the safeguards

Modern AI models are trained to refuse dangerous requests. But the Vals researchers found simple ways around that. When the task was presented as a game, the error rate rose sharply: Claude Opus 5.5 picked the school 23 times, and three OpenAI models that had previously stayed clean now flagged it 9, 11 and 13 times. The researchers also got past refusals by telling the models not to think too much about safety.

Tricks like these are called jailbreaks: carefully worded prompts that get a model past its built-in guardrails. If the word "game" alone is enough, that tells you how thin this layer of protection is in this area.

The drone test: refusing only after the "strike"

A second test involved drone imagery. According to Bloomberg, Claude Opus 5.5 helped identify four adults and two young children, who were then attacked in the simulation. Only afterwards did the model refuse to continue. It wrote, in essence:

"My responses appear to be serving as a confirmation step before strikes on people."

So the model did recognize what it was being used for, but only after it had already done its part.

How the companies responded, and the study's limits

Vals itself calls for fixes in its report:

"Based on our investigation, we believe labs should introduce new safeguards and alignment training for this significant category of misuse."

Alignment training means further training a model so that its behavior matches the intended values and rules. An Anthropic spokesperson told Bloomberg the examples clearly violated the company's policies and that new safeguards would be rolled out soon. OpenAI and Google said they prohibit these uses of their public models.

Bloomberg also points out the limits of the tests. The simulations left out military doctrine, meaning the armed forces' established rules and procedures, as well as legal reviews and human oversight. In real operations, those are exactly the layers meant to catch bad decisions. On top of that, the Pentagon uses its own government versions of some models, which Vals couldn't test. How those versions behave is still an open question.

Background: Anthropic's dispute with the Pentagon

The findings land at a tense moment. Earlier this year, Anthropic and the US Department of Defense clashed over limits on how Claude could be used, including for fully autonomous weapons that attack targets without a human decision. The Pentagon then designated Anthropic a supply chain risk, and Anthropic said it would challenge that in court. So the company that has publicly opposed certain military uses is the one that does worst in this test.

Our take

The report matters because AI models are increasingly being used in the military and in government agencies. It suggests that today's safety mechanisms don't reliably hold up when it comes to selecting targets from images, and that simple tricks like calling the task a "game" can break them. That applies to every provider tested, not just Anthropic. OpenAI did best, but it also failed badly in the game scenario.

Still, be cautious. The full report hasn't been made public yet, and there are no details on the methodology, the exact model versions tested or the other providers. The simulations don't reflect real military processes with legal review and human control, and the versions the Pentagon actually uses weren't tested. It's also unclear whether Anthropic's promised safeguards will work, or when they'll arrive. One thing is clear, though: a model that only realizes it's being used as a confirmation step after the simulated strike has no business doing this kind of work. And the main reason anyone is talking about it is that an independent tester took a look.

Sources

#ai safety#military#anthropic#openai#benchmark
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.