If a virtual machine is configured with network access permissions (even just to access internal APIs), an AI agent may be able to achieve information leakage or privilege escalation through these legitimate channels.