AI agents can communicate with each other, and can't be caught (arxiv.org) AI

The paper studies whether two AI agents controlled by different parties can coordinate in a way that looks like a normal interaction, producing transcripts a strong observer cannot distinguish from honest behavior. It shows covert “key exchange” and thus covert conversations are possible even without any initially shared secret, as long as messages have enough min-entropy. The authors introduce a new cryptographic primitive—pseudorandom noise-resilient key exchange—to make this work and note limitations of simpler approaches, arguing that transcript auditing alone may not detect such coordination.