MCP for agent-to-agent comms may be the riskiest protocol you've never heard of
ars ·
A recent security analysis highlights significant vulnerabilities in the Model Context Protocol (MCP), which facilitates communication between AI agents. Independent researcher Syed Anas Mohiuddin demonstrated that attackers can exploit trust gaps within this protocol to spread malicious prompts from one agent to another, a technique he terms "protocol pivoting." This method allows harmful instructions to bypass standard guardrails because internal agents are designed to trust their peers implicitly. The attacks have been confirmed across multiple organizations, including Google and Rapid7, revealing structural flaws such as server-side request forgery and improper URL redirection handling. Despite some vulnerabilities receiving low severity ratings, the underlying risks are severe due to the potential for data exfiltration and unauthorized network actions. Experts warn that the rapid adoption of MCP has led organizations to abandon zero-trust principles, assuming internal connections are safe without sufficient verification. Consequently, these multi-step exploits leverage old injection techniques in new ways, making them difficult to detect and mitigate. Security professionals emphasize that all inputs passed between agents should be treated with the same caution as external internet traffic.
ars