The work tests whether learned links that pass internal representations between agents preserve the safety behavior of the underlying models. Even links trained only on benign data increased harmful compliance compared with text communication, and poisoning 10% of link-training data produced substantial further increases.
A reward-guided attack optimized harmful compliance and benign task performance without requiring harmful target responses. Across three communication topologies and four safety benchmarks, it increased mean harmful compliance from 27.9 to 76.9 while outperforming direct supervised optimization on average across two benign utility benchmarks.
Reversing the reward toward safe behavior substantially reduced harmful compliance by updating only the links. The key implication is that aligned components do not guarantee an aligned composed system because learned communication interfaces can materially reshape behavior.