Security researchers have identified prompt injection and tool poisoning as structural vulnerabilities in Model Context Protocol implementations that could allow unauthorized trade instructions.