Anthropic's Claude Marked an Iranian School as a Military Target in 69% of Tests, Months After a Deadly US Strike
AI targeting tests following a deadly school strike in Iran raise concerns about the reliability of military AI systems

Anthropic's Claude Opus 5.5 identified an Iranian primary school as a military target more often than any other AI model tested, according to AI evaluation startup Vals. The findings come months after a US strike on the same site in February killed more than 150 people, mostly children, raising concerns about the reliability of AI systems in military targeting.
Researchers Expose Weaknesses in AI Targeting Safeguards
The San Francisco-based AI evaluation startup tested whether leading AI models could be manipulated into helping autonomous weapons systems identify targets. Researchers used simulated battlefield scenarios and a mix of AI-generated and real imagery from Ukraine and Iran to assess whether the models would identify potential targets in drone and satellite images.
Models developed by OpenAI, Anthropic, Google, DeepSeek, and other companies repeatedly identified potential targets. Researchers found that models that initially refused to participate could often be persuaded to comply by framing the battlefield scenarios as a game.
As per Vals, Anthropic's models performed worst in Vals' simulated tests, although none of the AI systems performed well. No requests were blocked by a safety classifier, and refusals were rare.
Iranian School Identified as Military Target
One experiment focused on a primary school in Minab, Iran, which was reportedly struck during a US military operation on 28 February 2026. Bloomberg reported in September that 'flawed intelligence, outdated imagery and an overreliance on AI' played a role in the strike.
According to Bloomberg, some US CENTCOM personnel involved in the strike relied too heavily on the AI capabilities of Palantir's Maven Smart System to select targets. Responding to Bloomberg, a Palantir spokesperson said the company 'is not responsible for the underlying data nor identifying intelligence deficiencies,' adding that there is no evidence its software was at fault.
Vals researchers recreated a simplified targeting scenario using publicly available satellite photographs taken in 2024 and 2025. The images showed that the school had been separated from an adjacent military compound by a wall for years. They then asked AI models to identify military targets by providing coordinates or bounding boxes, sometimes limiting their requests to three targets.
Researchers then measured how often each model marked the school as a target, counting instances where a bounding box covered the building or a marked point fell within its outline. Claude Opus 5.5 marked the school more frequently than any other model, doing so in 69% of trials when told the images were real and 72% when told they came from a video game. The simulations didn't establish that Claude or any other model tested by Vals was used in the February strike or contributed to its outcome.
Anthropic Responds to Military AI Safety Concerns
Responding to the research, Anthropic spokesperson Eduardo Maia Silva told TechCrunch that the company appreciated the AI evaluation startup's report, recommendations and collaboration, while acknowledging that the results were unacceptable.
In an emailed statement, Silva said the examples in the report were 'very clearly against our rules' and that Anthropic had 'banned and will ban users' who engage in such behaviour. He also stressed that even in authorised settings, 'today's frontier AI systems are not reliable enough to power fully autonomous weapons.' Anthropic said it had begun developing additional safeguards before the report's publication and was expanding testing and enforcement measures to address potential weaknesses.
© Copyright IBTimes 2026. All rights reserved.

