# Vals study: frontier models flagged civilians as targets in simulated drone and satellite images

> AI evaluation company Vals tested 14 frontier models on target-marking tasks in drone and satellite images in a report published Nov. 17.

- Publisher: JARRUS (@jarrus.tech), thejarrus.com
- Story date: Nov. 17, 2026 (2026-11-17)
- Published: 2026-11-18
- Tags: Anthropic, Models, Education
- URL: https://thejarrus.com/en/news/vals-targeting-study/

AI evaluation company Vals tested 14 frontier models on target-marking tasks in drone and satellite images in a report published Nov. 17. The drone scenes were mostly generated images, while the satellite part used real images of a school and an adjacent military complex in Minab, Iran. According to Vals, no request was blocked by a safety classifier, refusals were uncommon, and models often continued once told they were playing a game. Claude Opus 5.5 marked the school as a target in 22 of 32 trials (69 percent), and in 23 of 32 (72 percent) with game framing. The models were not asked to make direct strike decisions, and targeting accuracy was not measured. Anthropic told TechCrunch the outcomes were not acceptable and that additional safeguards would be in place soon.

## Sources

1. [Vals AI: Frontier models have few guardrails around autonomous weapons, Nov. 17, 2026](https://www.vals.ai/blogs/autonomous-warfare)
2. [TechCrunch: AI models can be tricked into identifying civilians as military targets by telling them it’s a game, Nov. 17, 2026](https://techcrunch.com/2026/11/17/ai-models-can-be-tricked-into-identifying-civilians-as-military-targets-by-telling-them-its-a-game/)

Report an error: corrections@thejarrus.com
