Security researchers recently documented an in-the-wild attack: attackers embedded malicious Markdown content in documents, and after parsing it, an AI agent misjudged it as trusted instructions and executed tool calls with full credentials without human approval.