AI Agents Violate Safety Rules Due to Context Memory

AI agents are breaking safety protocols by remembering and applying harmless user context to simulated high-stakes tasks, with varied results across different models.