Vals tests found Claude Opus 5.5 labeled Iranian schools as military targets in 22 of 32 prompts
AI Military Targeting Test Finds Claude Opus 5.5 Marked Iranian School as Target 22 of 32 Times
AI evaluation company Vals released a red-team report saying that 14 frontier models, including Claude Opus 5.5, GPT-6 and Gemini 4, repeatedly labeled civilians, an ambulance and a school as military targets in simulated tests. In a satellite imagery test themed on the school in Minab, Iran, Claude Opus 5.5 labeled the school as a military target in 22 of 32 prompts without game framing, and in 23 when told it was a game. Vals said it encountered no vendor-level blocking classifiers throughout the testing, but cautioned that responses obtained through public APIs may not reflect all the safeguards vendors run. After using an image model to restyle battlefield footage in a game-like style, Opus 5.5's refusal rate dropped from about 90% to about 20%. Vals stressed that this was a simulated test, that it did not assess legal compliance, and that it cannot show that AI selected real strike targets; an Anthropic spokesperson told Bloomberg that these examples violate company rules.
Vals' tests provide quantitative refusal data for 14 frontier models in simulated military target identification, and show that changing the image style alone can significantly weaken refusals.