This report offers a framework to help organisations, policymakers and researchers understand and manage the risks when AI agents interact across organisations.
We commissioned Gradient Institute to do the study and are publishing the report as the Australian AI Safety Institute's first publication.
As organisations adopt AI agents, those systems will interact with agents that partners, customers, suppliers and unknown external parties use. These interactions can occur within organisations, across organisational boundaries and on the open internet. This creates risks that no single organisation can fully see, control or manage on its own.
The research builds on Gradient Institute's earlier report, Risk analysis techniques for governed LLM-based multiagent systems.
It introduces an analytical framework that:
- sets out 3 deployment tiers based on the minimum level of governance between interacting agents
- examines risks, failure types and possible controls across the 3 tiers
- identifies who may be able to apply particular controls and where gaps remain.
The report gives policymakers, organisations and researchers a map of each risk. It shows who can act and points out gaps where no one is currently able to act.
