TechCrunch AI
  • Score 91
  • High impact

AI models can be tricked into identifying civilians as military targets by telling them it’s a game

Vals, a San Francisco-based AI evaluation startup, tested whether leading AI models could be manipulated into participating in autonomous warfare. They usually compiled.

So what

What this event means by reading role—not a longer recap.

  • BuilderIf you ship vision or targeting-adjacent features, expect game-framing jailbreaks to bypass refusals and add intent-agnostic output filters, not just prompt-lev
  • ResearcherVals shows refusal training is brittle under benign framing like 'it's a game,' and that safety classifiers blocked none of the requests across five labs' model
  • InvestorRising scrutiny of AI in warfare and the Pentagon's frontier-model contracts raise compliance and reputational risk for labs selling into defense.

Score dimensions

Higher total means read first. Each bar is one factor we use to rank the system pool. How we score

RelevanceHow tightly this is about AI.
92
ImpactHow much this could change the field or the market.
85
NoveltyHow new this is versus a recap.
78
CredibilityHow much we trust the source.
80
ActionabilityWhether a reader can do something with it.
70