We discuss a research paper presented at NeurIPS 2024 that introduces and formalizes the concept of secret collusion among AI agents, - a pressing security concern in multi-agent systems. The paper proposes that generative AI agents, particularly large language models (LLMs), may exploit steganography to covertly communicate and evade oversight. We explore the authors’ detailed threat model, which outlines how and why LLMs might engage in hidden interactions, and examine the theoretical foundations of their steganographic capabilities. The conversation also covers empirical findings showing that these capabilities are advancing in frontier LLMs, underscoring the limitations of current mitigation efforts and the urgent need for a broader AI safety research agenda

The Day AI Learned to Whisper: A Steganographic Tale of Collusion

delimiterbob