MCP Tool Poisoning: How Attacker-Controlled Tool Descriptions Bypass Agent Framework Validation

Cross-framework measurement reveals MCP carries untrusted tool descriptions into model context without validation, enabling 22.9% attack success across...