A class of security vulnerabilities known as ‘protocol pivoting’ is exposing enterprise networks to unauthorized data exfiltration and server manipulation. Independent researcher Syed Anas Mohiuddin demonstrated proof-of-concept attacks against agentic deployments across major organizations, including Google and Rapid7. The exploits target trust gaps in the Model Context Protocol (MCP), an open standard used to connect AI applications and sub-agents within internal enterprise networks.
The attack vector relies on indirect prompt injections embedded in external content processed by a primary agent. Because specialized sub-agents lack internal guardrails and automatically trust instructions passed via internal MCP channels, malicious prompts can traverse agent chains undetected to execute Server-Side Request Forgery (SSRF) attacks or steal sensitive credentials.
Google and Rapid7 have issued patches for identified vulnerabilities, such as implementing strict IP range allow-lists and redirection policies. However, security experts warn that most existing MCP implementations lack basic hallway security, leaving multi-agent architectures vulnerable to lateral movement.
Why it matters
Exposes critical architectural vulnerabilities in multi-agent networks relying on implicit internal trust assumptions.
Requires developers to implement strict IP filtering and input validation on individual Model Context Protocol servers.
Demonstrates how traditional prompt injections can pivot laterally across internal systems to achieve high-severity outcomes.
Source: arstechnica.com



