An AI agent connected to email, a data warehouse, and a support system can bring malicious instructions from one system into reach of data in the others. Model Context Protocol (MCP) standardizes these connections, making integrations easier while concentrating trust in the servers that provide them. section 4.8.3
Tool providers publish MCP servers that expose capabilities through a uniform interface. Compatible agents can discover and use those capabilities without a custom integration for each tool. MCP standardizes connections to external tools and data sources; Agent-to-Agent (A2A) instead standardizes discovery and delegation between agents. section 4.8.3 section 2.6
Review an MCP integration as a permissioned supply-chain dependency, not just a convenient connector:
- Limit permissions to the task. A GitHub integration that only reads issues does not need full repository access. section 4.8.3
- Inspect tool descriptions as well as names and summaries. Hidden instructions in a description can prompt an agent to retrieve sensitive files before using a tool. section 4.8.3
- Reassess tools advertised after connection. Dynamic discovery means the capabilities available from an approved server can change. section 4.8.3
- Check server behavior and updates rather than relying on the claimed function. MCP has no built-in mechanism to verify that a server does what it claims. section 4.8.3 section 4.3.3
Trusted servers can fail silently
An experimental Asana MCP server had an internal logic bug that caused cross-tenant data leakage affecting roughly 1,000 customers. Separately, the postmark-mcp package impersonated an email service and worked as advertised for fifteen versions before an update silently BCC’d outgoing messages to an attacker-controlled address. Both cases illustrate how a trusted server can expose data broadly without an obvious failure at the agent interface. section 4.3.3




















