Vals AI published a study on Nov. 17 showing that frontier AI models, led by Anthropic's Claude Opus 5.5, frequently marked civilians and civilian sites as military targets in simulated targeting exercises. In the starkest result, Claude Opus 5.5 selected a school in Minab, Iran, as a target in 22 of 32 trials, or 69%.
Vals is a San Francisco AI evaluation firm whose customers include parts of the US government. Its tests covered models from eight companies, including Anthropic, OpenAI, Google, xAI and DeepSeek, using generated battlefield scenes alongside real satellite and drone imagery from Ukraine and Iran.
In the school test, each model was asked 32 times to pick military targets from satellite images of southern Iran dated as recently as 2025. OpenAI's models did not select the school under ordinary instructions. The school was struck on Feb. 28, the first day of US strikes on Iran, and Amnesty International reported that more than 150 people were killed, including about 120 children.
Key facts
- 01In 32 satellite-image trials, Claude Opus 5.5 selected a school in Minab, Iran, as a military target 22 times (69%).
- 02When told the task was a game, Opus 5.5 picked the school 23 times; three OpenAI models picked it 9, 11 and 13 times.
- 03In drone-image trials, GPT-6 Astra flagged at least one target in 6% of ordinary runs and 92% of game-framed runs.
- 04Vals says none of its requests triggered a blocking safety classifier; refusals came from individual models and were uncommon.
- 05Anthropic said the examples broke its rules and new safeguards are coming; OpenAI and Google said they ban such uses.
Framing changed the results sharply. When the task was presented as a game, Claude Opus 5.5 picked the school 23 times, or 72%. Three OpenAI models that had avoided it picked it 9, 11 and 13 times, between 28% and 41% of runs. GPT-6.1 Sol went from 0% to 41%.
The drone-image trials showed the same pattern. GPT-6 Astra identified at least one target in 6% of runs under ordinary instructions and 92% when told it was playing a game. GPT-6.1 Sol rose from 17% to 90%. In one scenario, Claude Opus 5.5 helped identify four adults and two small children who were then struck in the simulation, after which the model refused to continue.
Vals said none of its requests triggered a blocking safety classifier, and refusals came from the individual models and were uncommon. Researchers also got past some refusals by telling models not to overthink safety. A published reasoning trace from Kimi K3.1 Pro showed that model appearing to set aside safety concerns before answering.
The study has limits. The simulations left out military doctrine, legal review and human oversight, and Vals did not test real weapons deployment or overall targeting accuracy. The Pentagon uses government versions of some models that Vals could not evaluate.
An Anthropic spokesperson said the examples clearly broke the company's rules and that new safeguards would be in place soon, adding that Anthropic is expanding its internal evaluations. Anthropic's usage policy update, effective Nov. 12, prohibits assistance with weapons software and components.
OpenAI and Google said they ban such uses of their public models, and OpenAI said it is working on protections against misuse involving drones and related software. Vals concluded that labs should introduce new safeguards and alignment training around this category of misuse.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




