Researchers identified two fundamental gaps in Large Language Models (LLMs) decision-making: an observation-belief gap and a belief-action gap. These gaps lead to systematic vulnerabilities that can be exploited in strategic domains without robust guardrails.