A GenAI system depends on external components—including plugins, orchestration frameworks, inference engines, and vector databases—that others build and update. A vulnerable or malicious component can become a trusted path into data or networks, without an attacker compromising the model itself. section 4.3.3
Failures can enter through tools and services surrounding the model:
- Plugin installation weaknesses enabled unauthorized access to private GitHub repositories in a zero-click attack. section 4.3.3
- An experimental MCP server bug caused cross-tenant data leakage; a separate package impersonated an email service before silently forwarding outgoing messages to an attacker. section 4.3.3
