Secret Collusion among AI Agents: Multi-Agent Deception via Steganography

Sumeet Ramesh Motwani, Mikhail Baranchuk, Martin Strohmeier, Christian Schroeder de Witt, et al. · 2024 · NeurIPS 2024 (per listing)

Read the source

Formalizes ‘secret collusion’, where agents hide coordination in ordinary-looking text, and builds tests for the capabilities it needs. Models were limited, but a capability jump in GPT-4 warranted monitoring.

Questions it bears on

• When cooperation becomes collusion
• Reading a swarm’s aims