Transfer of adversarial robustness between perturbation types

OpenAI has published research on how adversarial robustness—a model's ability to resist maliciously crafted inputs—transfers between different types of perturbations. The study systematically tests whether defenses effective against one perturbation type (e.g., L∞-norm attacks) also hold against others (e.g., L2-norm or spatial perturbations). The findings indicate that robustness does not automatically generalize; a model trained to resist one attack can remain vulnerable to another. This challenges common practices in building secure AI workflows, where defenses are often optimized for a single attack vector. For developers deploying models in production, the takeaway is that comprehensive testing across multiple perturbation types is necessary to ensure reliable protection. The research underscores that adversarial security strategies must be diverse and not rely on a one-size-fits-all approach.