Most AI security testing stops at the chat box. Malicious model files, poisoned datasets and MCP tool abuse — the ML supply chain attack surface, tested.