Secret Collusion Among Generative AI Agents: Toward Multi-Agent Security
With Christian Schroeder de Witt
I argue that the ability of advanced AI agents to use perfect stealth will soon be AI Safety’s biggest concern.
Importantly, information-theoretic undetectability undermines our current approaches to multi-agent security.In this talk, I will be focusing on the matter of steganographic collusion among generative AI agents. I will give an introduction to steganography, talk about how agents could be incentivised to use it in unintended ways, and present a novel model evaluation framework, as well as mitigation measures. How do we design autonomous systems and environments in which undetectable actions cannot cause unacceptable damages?