Safety of Latent Communication in Multi-Agent Systems

[2026.10] Paper released on arXiv. Code is being refactored and will be released soon.

We show that learned latent communication links can compromise the safety of multi-agent systems built from safety-aligned agents. Even benign link training can increase harmful compliance, and attacks via supervised optimization, data poisoning, or reward-guided optimization amplify it further. Reward-guided updates to the links alone can repair compromised systems.

@article{huzaifa2026safety,
  title   = {Safety of Latent Communication in Multi-Agent Systems},
  author  = {Huzaifa, Muhammad and Mavali, Sina and Eisenhofer, Thorsten},
  journal = {arXiv preprint arXiv:2609.39788},
  year    = {2026}
}

⭐ Star or watch this repository to be notified when the code is released.