research

Trading inference-time compute for adversarial robustness

For developers building AI workflows, this trade-off offers a potential new way to harden systems against adversarial attacks without retraining, but at the cost of increased computational overhead per request.

OpenAI Blog·January 22, 2025·1 min readresearch

researchTrading inference-time compute for adversarial robustness

openai.com

What happened

OpenAI has published a blog post exploring the idea of trading inference-time compute for adversarial robustness. Instead of relying solely on model architecture or fine-tuning, the post suggests that allocating additional computational resources during inference can help defend against adversarial inputs—subtle perturbations designed to fool AI models. This approach treats robustness as a runtime investment, where more compute per query can yield stronger resistance to malicious manipulation. For builders of AI workflows, this introduces a new dimension to consider: the balance between inference speed/cost and security. While not a fully realized product, the concept aligns with broader trends in using test-time compute scaling to improve model performance and safety. The practical angle lies in evaluating compute budgets against potential risks in production systems, especially those handling untrusted user inputs.

Key takeaways

OpenAI discusses using extra computation during inference to improve adversarial robustness, rather than only through training or architecture changes.
The approach treats robustness as a per-query cost that can be scaled dynamically based on threat level.
Builders may need to factor inference-time compute into their security planning for AI applications.
The concept is part of a larger research area exploring test-time compute as a lever for performance and safety.