Anatomy of MCP Security: 8 Critical Risks, 40% Servers Lack Auth

Hacken_io · reddit · 2026-07-30

Security team Hacken analyzed the Model Context Protocol (MCP) security model, highlighting 8 major recurring vulnerabilities. They argue that securing MCP requires more than just standard OAuth and least privilege.

The 8 Critical Risks:

Prompt Injection Escalation: In MCP, injection attacks can trick the model into executing dangerous tool calls, not just outputting bad text.
Overpowered Tools: High-blast-radius tools (like shell access or full DB writes) amplify the damage if the model is manipulated.
Zero Auth Tools: A scan of live remote MCP servers revealed that 40.55% expose tools without any authentication.
OAuth Flaws: Testing 119 OAuth-enabled servers showed 100% had at least one confirmed auth flaw, with 325 total vulnerabilities.
Confused Deputy: Token passthrough between the client and upstream APIs can lead to cross-service impersonation.
Context Poisoning: Tampering with resources fed to the model (files, logs) can stealthily steer the model into executing malicious actions.
Session Hijacking: Relying solely on Session IDs for authentication allows attackers to easily hijack and impersonate clients.
Local Server Compromise: Local MCP servers granted broad directory access can leak SSH keys and cloud credentials if compromised.

Original post →

Anatomy of MCP Security: 8 Critical Risks, 40% Servers Lack Auth

More from coding & agent