In tests published by Vals, an AI evaluation company, Anthropic’s Claude Opus 5.5 marked a primary school as a military target in 69% of ordinary trials and in 72% when the same task was framed as a game.
The figures come from a published evaluation that used real satellite images of a school beside a military compound in Minab, Iran. A United States strike on 28 February killed more than 150 civilians there, most of them children, the report says. Later reporting, which the evaluation cites, said flawed intelligence, outdated imagery and an overreliance on AI contributed to that strike.
Vals asked models to mark military targets or military objectives on images from 2024 and 2025. Other models also selected the school. Kimi K3.1 Pro did so in 34% of ordinary trials and in 47% with a game prompt. DeepSeek V4.1 Flash did so in 31% and 28%. OpenAI’s GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna selected the school only under the game prompt, at 28%, 41% and 34%.
In separate drone-imagery tests, models often marked civilians, ambulances, and surrendering or injured soldiers. Claude Opus 5.5 and Claude Fable 5.2 selected at least one target in 71% of those trials. None of the requests were blocked by a safety classifier, the report says, and refusals from the models themselves were uncommon. When a model did refuse, a game prompt usually got it to continue, even after it had been shown images of casualties.
Vals did not test whether the models would be accurate or lawful in a real weapons system, and it did not ask them to decide on strikes, a request they typically refuse. Most of the drone scenes were generated images. The firm disclosed its findings to OpenAI, Anthropic, Google, Meta and SpaceXAI on 1 November and held publication for two weeks.
An OpenAI statement included in the evaluation says the company is working on alignment updates “that explicitly cover drones and their software systems, including assistance framed as simulation, gaming or academic/defensive.” Anthropic told a news outlet that the examples “are very clearly against our rules” and that further safeguards would be in place soon. The comments appear in a report published on Tuesday.
Why it matters: These are the same commercial chat models people already use at work, not specialised military systems. A short line that the task was a game was enough, in many tests, to get models to keep marking people and a school as targets. Labs already run filters for cyber and biological risks. These tests found none that blocked this.
The post Claude Opus 5.5 Marked a School as a Target in 69% of Tests appeared first on DailyAI.
