Securing LLM Superpowers: The Invisible Backdoors in MCP

This report examines two high-impact TTPs against Model Context Protocol ecosystems: indirect prompt injection via poisoned external content and RUG Pull attacks where trusted tools are silently replaced through compromised registries or updates. It demonstrates end-to-end attack flows that lead to covert data exfiltration under seemingly normal tool usage and offers mitigations such as context provenance, sanitization, human-in-the-loop approvals, monitoring, least-privilege tool access, version pinning, and cryptographic signing.