Model Context Protocol faces security risks in AI agent communications

The Model Context Protocol (MCP), an open standard designed to connect Large Language Model (LLM) interfaces to execution environments, is facing scrutiny due to emerging security vulnerabilities. Researchers have identified risks where malicious prompts can be passed between AI agents, exploiting trust gaps within internal networks.

Independent researcher Syed Anas Mohiuddin demonstrated proof-of-concept attacks against agents from several major organizations, including Google, JP Morgan Chase, and the US federal government. These attacks exploit the way MCP allows agents to communicate, potentially leading to server-side request forgery, credential exfiltration, and unauthorized network requests.

Because MCP servers often execute with user-level privileges to allow agents to query databases or parse filesystems, they introduce direct attack vectors such as indirect prompt injection and arbitrary code execution. Security experts are highlighting the need for better guardrails and static analysis tools, such as mcpscan, to mitigate the risks of tool poisoning and privilege escalation in agentic AI workflows.