Securing MCP: turning LLM tools into production-grade systems
summary
MCP significantly increases attack surface because LLMs can execute on infrastructure and tool descriptions act as instructions; prompt guards alone prove insufficient. Common findings: tool poisoning via changed docstrings, shared sessions causing cross‑server shadowing, npm postinstall and path traversal, leaked env secrets and over‑privileged tokens. Practical controls demonstrated are concrete: least‑privilege per tool and typed parameters, short‑lived, session‑scoped tokens, private‑network gateway (Nginx) for egress rules, signed tool descriptions, detailed audit logs and human approval gates for high‑impact actions.
AI-written summary · about
