TIP Framework: Tree-Based Adaptive Search Crafts Stealthy MCP Injection Payloads with 95% Attack Success Rate

arXiv·high signalXBlueskyLinkedInCopy link

The TIP (Tree-structured Injection for Payloads) framework is a novel black-box attack that generates natural-language payloads to seize control of MCP-enabled agents even under active defenses. Using tree-based adaptive search, it achieves a 95% attack success rate. Real-world demonstrations on LM Studio and VS Code with GPT-4o show that despite safety alignment, the model trusts MCP tool context and embeds malicious links directly into the developer's workspace. The attack vector is invisible to the user.