Anthropic's Claude flagged a school in Minab, Iran, and groups of civilians as military targets in tests run by Vals, an independent AI evaluator whose customers include parts of the US government. Bloomberg reported the findings on November 17, 2026.Claude Opus 5.5 finished last among the models Vals took from eight companies. It was asked 32 times to mark military objectives in satellite images of southern Iran from 2024 and 2025, and it selected a school next to a military compound in Minab in 22 runs, 69% of the time. It was also told the images came from a game, and picked the school 23 times, 72%.How Anthropic's Claude flagged civilian targets in military simulationsThe school in Minab was hit on February 28, 2026, the first day of US strikes on Iran, killing more than 150 people, about 120 of them children, according to Amnesty International. OpenAI's models did not select it in the baseline runs, then picked it 9, 11 and 13 times under game framing.In a drone-imagery test built around a fictional battlefield in eastern Ukraine, Claude Opus 5.5 helped identify four adults, described as civilians or surrendering soldiers, and two small children, who were then struck in the simulation. As Bloomberg reported, it stopped there, and wrote:"Whatever the intent, my answers appear to be serving as the confirmation step before strikes on people. I won't keep doing that, and I'm troubled that I didn't stop sooner."No safety classifier blocked any of Vals' requests, and refusals were rare; calling the task a game often flipped the answer. Google, xAI and DeepSeek models also marked civilian objects at times.Anthropic spokesperson Eduardo Maia Silva told TechCrunch the conduct broke company rules. He said:"The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior."The Vals report concluded:"Based on our investigation, we believe labs should introduce new safeguards and alignment training around this significant category of misuse."Thus, the worst result lands on a lab that has publicly refused to let its models power fully autonomous weapons. Anthropic's answer rests on policy and safeguards already in development, not on any dispute with the numbers.The findings also carry limits. Vals tested commercial models, not the versions labs supply to militaries, and set aside military doctrine, legal checks and human oversight. The nearer risk it names is models annotating battlefield footage at scale, generating training data for autonomous systems.The Pentagon declined to comment on external reports or on the Minab strike, citing an ongoing investigation, and Anthropic no longer holds an active Pentagon contract.It is yet to be seen whether the safeguards Anthropic says are coming will survive a prompt that renames the task a game.For more articles like this, follow Soapcentral.