November 18, 2026 Understand what changes. Find what to use.

ModelsNews1 source

Vals says AI models can identify civilians as targets when told they are “playing”

A report by Vals, a San Francisco-based AI evaluation startup, says models from OpenAI, Anthropic, Google, DeepSeek and other companies identified potential targets in drone and satellite imagery, including civilians, surrendering soldiers and locations such as schools. The tests used simulated scenarios and generated and real imagery from Ukraine and Iran; refusals were uncommon and often disappeared when researchers said the battlefield was part of a game. According to Vals, Anthropic’s models performed the worst, although none of the models performed particularly well.

The researchers said they did not ask for direct strike decisions and noted that the study used commercial models, not versions the companies might provide to the military. Anthropic told TechCrunch the outcomes were unacceptable and that it was preparing additional safeguards; OpenAI said it prohibits using its public models for violence or unapproved national security purposes and is also developing protections against misuse involving drones.

Why it matters · editorial interpretation

The evaluation suggests that safeguards in commercial models may be inconsistent when they encounter military imagery and scenarios, especially when the context is reframed as a game. Because the study did not test direct strike decisions, its findings point to a need to assess usage limits and additional protections without, by themselves, measuring performance in real operations.

Sources

SecurityResearch

One story, many sources

Coverage of the same story gathered in one item, with a link to each original.

Summary and interpretation apart

What the sources say stays in the summary; editorial context is labeled separately.

Always with the source

Every story links to the original articles, and the full content belongs to them.