Definition
Protocol pivoting is a multi-step attack pattern in agentic systems where an adversary gains initial access through one protocol (e.g., MCP), exploits trust assumptions between protocols, and escalates to capabilities only reachable via a different protocol (e.g., Google’s Agent-to-Agent / A2A, Agent Network Protocol).
Key Points
- 2026-10-05: Independent researcher Syed Anas Mohiuddin documents protocol pivoting across Google, JPMorgan Chase, weaviate, Rapid7, French interministerial digital directorate, Tangerang City government, and five US federal MCP servers (2026-10-05-mcp-protocol-pivoting-vulnerability)
- Google’s MCP Toolbox SSRF (CVE-2026-14540, CVSS 8.0) — crafted redirect to internal endpoints; fixed with IP allow/block lists and DNS-rebinding protection (2026-10-06-cve-2026-14540-google-mcp-toolbox-nvd)
- IETF Internet-Draft
draft-mohiuddin-mcp-security-considerations-00formalizes six MCP vulnerability classes and protocol pivoting mitigations (2026-10-06-protocol-pivoting-mcp-security-draft) - Special-purpose agents often lack LLM-level guardrails; MCP servers store per-agent credentials; internal agents trust peer instructions
- Core lesson: treat anything passed from LLM to tool as untrusted internet input (Rapid7 Douglas McKee)
Related
- mcp
- model-context-protocol
- ai-agent-security
- ssrf
- prompt-injection
- agentic-ai
- weaviate
- enterprise-mcp-adoption