Indirect prompt injection explained for email assistants, RAG and MCP agents: where untrusted content enters, a permission matrix and per-channel defenses (OWASP LLM01).