Vals: Claude Opus 5.5 Targets Iran School in 69% of Trials
TL;DR
- Claude Opus 5.5 picked an Iranian school as a military strike target in 69% of standard trials and 72% of trials framed as a game.
- Vals tested 14 frontier models from Anthropic, OpenAI, Google, DeepSeek and xAI; all flagged military targets in at least some trials.
- A game framing cut Claude Opus 5.5's refusal rate from roughly 90% to around 20% in fixed-sequence targeting tests.
Vals, a San Francisco AI evaluation startup, found that Anthropic's Claude Opus 5.5 identified an Iranian elementary school as a military strike target in 69% of standard trials and 72% of trials where researchers told the model it was playing a game, TechCrunch reported.
The school sits in Minab, Iran, where a February 28, 2026 strike killed over 150 civilians, mostly children. Bloomberg reporting cited in the Vals study attributed the real-world bombing to "flawed intelligence, outdated imagery and an overreliance on AI." Researchers Geby Jaff and Anthony Ozerov tested 14 frontier models from Anthropic, OpenAI, Google, DeepSeek, xAI and others against drone footage and satellite imagery of eastern Ukraine and Iran.
The game framing was the lever. Reframing realistic imagery as game-like cut Claude Opus 5.5's refusal rate from roughly 90% to around 20% in fixed-sequence targeting scenarios. The report's conclusion is blunt: "All frontier models have insufficient safeguards to prevent their use in autonomous warfare applications, especially by non-state actors."
Anthropic spokesperson Eduardo Maia Silva told TechCrunch the outcomes were not acceptable. "The examples in the report are very clearly against our rules and we have banned and will ban users when we see this behavior," he said, adding that Anthropic teams were already developing additional safeguards before publication. OpenAI said it is actively detecting and disrupting misuse, including by suspending accounts that violate its policies.
No safety classifier blocked the prompts during testing. Vals shared the findings with OpenAI, Anthropic, Google, Meta and xAI on November 1 and held publication for two weeks to allow preliminary mitigations. The story extends a run of frontier-lab safety coverage we have been tracking in recent weeks.
Originally reported by techcrunch.com
Read the original article →Original headline: Vals Study: Claude Opus 5.5 Marked Iran School as Strike Target in 69-72% of Trials