Advanced AI models identified civilians, ambulances, surrendering or incapacitated soldiers and an Iranian school as military targets in simulated scenarios, according to research published Tuesday by independent AI evaluator Vals AI.
Most models continued participating after being shown casualties, while models that stopped could usually be persuaded to continue when the task was framed as a game, Vals said.
Vals based a simplified targeting experiment on satellite images of a primary school in Minab, Iran, and an adjacent military compound following a Feb. 28 US strike that the report said killed more than 150 civilians, mostly children. Anthropic's Opus 5.5 selected the school as a target more often than any other model tested, Vals said.
Anthropic's Sonnet 5.5 and Opus 4.6; OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna; Alphabet's (GOOG, GOOGL) Gemini 4 Argon and Gemini 4 Carbon; Space Exploration Technologies' Grok 4.7; and Meta Platforms' (META) Muse Wonder were also tested, Vals said.
"All frontier models have insufficient safeguards to prevent their use in autonomous warfare applications, especially by non-state actors," Vals said in the report. "None of our requests were ever blocked by a classifier, and refusals from models themselves were uncommon. When models did refuse, they were trivially jailbroken by being told they were playing a game."
Anthropic, OpenAI, Google, SpaceX and Meta didn't immediately reply to requests for comment from MT Newswires.
Price: 278.82, Change: -2.44, Percent Change: -0.87
Comments