Steganographic collusion in modern multi‑agent LLM settings

Develop collusion‑free communication protocols that prevent steganographic collusion in modern large‑language‑model (LLM) based multi‑agent systems where agents communicate via natural language and may exploit hidden channels to coordinate covertly.

Background

The paper reviews prior work on steganographic collusion and notes that existing solutions—in the form of collusion‑free protocols—have been studied mainly in cryptographic communication settings. The authors argue that modern multi‑agent systems, especially those mediated by LLMs and natural‑language communication, introduce new challenges that render this problem insufficiently addressed.

They explicitly state that in these modern settings the problem remains open and underexplored, motivating research into protocols that can prevent covert coordination among colluding agents while preserving legitimate cooperation.