Governance and security of connected artificial intelligence in Australian businesses

Abstract

Recent Australian security guidance identifies unauthorised access to artificial-intelligence services and the risks introduced when models connect to organisational tools. This paper examines the relationship between account security, indirect prompt injection and enforced operational controls. It reviews Australian Signals Directorate guidance, original prompt-injection research, an ACL benchmark and the NIST generative-AI risk profile. The study uses a targeted narrative review and a fictional business workflow to distinguish model behaviour from the permissions of the surrounding system. It finds that credential protection and behavioural safeguards address different failure pathways. A model instruction alone cannot demonstrate that sensitive actions are technically constrained, while a benchmark result cannot establish the safety of a particular deployment. The paper proposes controls and an evaluation plan for a small organisation, including bounded permissions and repeatable testing. It does not assess any live AusLume integration, estimate Australian incident prevalence or certify a product. The conclusion is that connected AI should be evaluated as an operational system, with accountability and demonstrable controls matched to the consequences of its actions.