Introduction
This report looks at the risks of multiple AI agents interacting with each other. Agents are AI systems that can perceive and take actions in its digital environment with some degree of autonomy.
We commissioned the Gradient Institute to do this study as part of our work to better understand emerging risks from AI for Australia.
The report found that when multiple agents are deployed within a single environment, this fundamentally changes the risks landscape. The report identified 6 key failure modes which could arise from multi-agent interactions. These result from a lack of validity, which is whether assessments measure what they claim to measure and produce reliable results that align with real-world outcomes.
The report equips organisations with foundational knowledge and tools for analysing the distinctive challenges emerging from multi-agent interactions in their governed environments.
This helps give organisations important practical guidance to effectively manage implementation risks when deploying AI technologies.