Transfer of adversarial robustness between perturbation types

In three linesOpenAI publishes research on transfer of adversarial robustness across perturbation types. The study demonstrates that models trained against one adversarial attack type develop partial robustness to other attack types, offering insights for improving AI system defenses.

Summary generated by Claude — human-verified