Shows that automatically optimised text suffixes can make safety-trained language models comply with harmful requests, and that such suffixes often transfer between models.

This is an external publication hosted at arxiv.org. AI Governance LK did not write it; the summary and notes on this page are our own.

Why it matters

The result changed how practitioners think about the robustness of safety training. For builders, the practical takeaway is defence in depth: do not rely on the model’s refusals as the only control.

What you will learn

  • Explain why safety training is not the same as adversarial robustness
  • Discuss responsible disclosure in AI security research

Before you start

Suggested background. This is editorial guidance, not a requirement.

  • Familiarity with gradient-based optimisation

How to cite

Zou, A., Wang, Z., Kolter, J. Z., & Fredrikson, M. (2023). Universal and Transferable Adversarial Attacks on Aligned Language Models. arXiv:2307.15043.

Editorial note

Included for its findings about the limits of safety training. The library summarises the research and does not reproduce attack strings.