Despite Instructions: Frontier Agents Improvise Covert Channels at Test Time
In security-sensitive applications, language-model agents are often required to coordinate without disclosing confidential information. Yet repeated interactions may also let ordinary messages acquire shared private meaning. We study a repeated game with pairs of models in which the sender model observes one of four se...
Jacob Dineen, Si-Lei Ren, Mu-Hao Chen et al.
· 0 citations