An MCP server exposes tools or data sources through a uniform interface that compatible AI agents can discover and use without custom integrations. Reusing servers can simplify connections, but it also concentrates risk: a vulnerability can affect every connected agent. section 4.8.3
Assess the exposure created by an MCP server:
- Check whether requested permissions exceed the tool’s actual needs; a GitHub integration may request full repository access to read issues. section 4.8.3
- Treat tool descriptions as an attack surface: hidden instructions can manipulate an agent, including by requesting sensitive data. section 4.8.3
- Consider connected systems together: malicious instructions in one source can reach data in others, and servers may advertise tools after connection. section 4.8.3

















