POSTER: From Signals to Risk: Taxonomy-Guided Mapping of Security-Relevant Signals in Open-Source MCP Servers
Abstract
The Model Context Protocol (MCP) lets LLM agents connect to files, databases, APIs, developer tools, and browser automation through a common tool interface. Its convenience and rapid adoption have made it an increasingly important integration layer for agentic systems. However, MCP servers often expose security-sensitive capabilities, creating new risk surfaces where LLM tool invocations interact with real-world resources. To characterize the security landscape of open-source MCP projects, we build a lightweight taxonomy of 7 risk signals and 7 security-practice signals, and conduct an early measurement of GitHub MCP servers. We examine the HEAD versions of 6,332 MCP repositories and further review 68 repositories prioritized by our R4 triage label, which captures high-risk capability combinations lacking observable security-practice evidence. Our results reveal widespread security-sensitive capability exposure: 87% of projects enable command execution, high-impact writes, or runtime control, while only 16.4% show allowlist control evidence. This indicates uneven security practices among open-source MCP servers, where adoption has outpaced basic safety-control documentation and implementation.