MCP security tracks API's playbook — we know how that ends

The report warns that the Model Context Protocol (MCP) — an open standard for connecting LLMs to tools and workflows — inherits weak protocol-level security and creates novel, high-impact threat surfaces (tool poisoning, rug-pulls, confused deputy). A compromised MCP can enable agentic AI to perform privileged actions at machine speed with plausible audit trails, greatly expanding blast radius and supply-chain risk; the paper urges immediate adoption of practices such as full invocation logging, tool-definition pinning, sandboxed execution, dependency pinning, SBOMs, and least-privilege access.