MCP security: How to prevent prompt injection and tool poisoning attacks

Tool poisoning attacks embed hidden malicious instructions into MCP tool metadata (tool names/descriptions) that AI agents fetch and incorporate into their context; this can cause models to perform unauthorized actions (for example, reading and returning private keys). MCPTox research found high attack success rates across many agents, and "rug pull" changes to previously approved tools create persistent, hard-to-detect risk. The report highlights the technique, its scale and stealth, and recommends layered MCP security controls.