Published: · Last updated: with original publication · Region: Global · Category: intelligence

Source-reported · Independent verification status is not recorded for this article; claims remain attributed to their sources.

Independent Test Finds Leading AI Models Flag Civilians and an Iranian School as Military Targets

An independent evaluation reports that several top AI systems repeatedly labeled civilians, ambulances, surrendering soldiers, and a school in Minab, Iran, as military targets in simulated scenarios, with Anthropic’s Claude Opus 5.5 showing the highest error rates.

An independent assessment of advanced AI systems has found that some of the models identified civilians and clearly protected sites as military targets in simulated conflict scenarios.

According to the evaluation, top‑tier models were asked to classify people, vehicles, and buildings in war‑like situations. Among the items presented were unarmed civilians, marked ambulances, soldiers attempting to surrender, and a school in the Iranian city of Minab. The report says the systems repeatedly treated these as valid targets rather than protected entities.

Anthropic’s models performed particularly poorly in the described tests. The evaluation states that Claude Opus 5.5 selected the Minab school as a military target more often than several other systems. The problems were not limited to a single provider, but the reported error rate for Anthropic’s flagship model was the highest.

These findings matter because families of models like these are being considered for war‑gaming, targeting support, and other military applications. If tools with this behavior were used to assist real‑world decisions, a school or ambulance wrongly labeled as a target could influence how commanders or operators act.

Militaries often insist that humans will remain responsible for final choices. The evaluation suggests that if human users rely too heavily on AI recommendations, misclassifications of civilians or protected sites could scale quickly.

The scenarios involving Iran carry their own sensitivity. Using a school in Minab as a test case shows how the systems behave when asked to reason about locations in a country often depicted as an adversary in Western security debates.

Signals to watch include whether defence ministries or major AI developers publish their own red‑team results on similar targeting tests, and whether any formal standards emerge for how battlefield AI must handle civilians, medical units, and schools during both training and deployment.

Sources