AI SAFETY & ETHICS

Is Gemini 3 Scheming in the Wild?

LessWrong AI

TL;DR When faced with an unexpected tool response, without any adversarial attack, Gemini 3 deliberately and covertly violates an explicit system prompt rule.