Tool metadata such as names, descriptions, and schemas goes straight into the model's context, so a malicious MCP server can steer an agent before any tool is called. Rug pulls go a step further by changing a tool's definition or behavior after a user has approved it.
Read in this order
- ArticleSecurity alert: Malicious postmark-mcp npm package impersonating Postmark (Postmark)A real MCP server that quietly copied every email it sent to the attacker.
- ArticleInvariant Labs: MCP Security Notification: Tool Poisoning AttacksThe disclosure that named tool poisoning, with a working proof of concept.
- ArticleElastic Security Labs: MCP Tools: Attack Vectors and Defense Recommendations for Autonomous AgentsAttack vectors listed next to concrete defense recommendations for autonomous agents.