Risk Under Pressure: Compute-Aware Evaluation of Adversarial Robustness in Language Models

In three linesEvaluation framework for LLM adversarial robustness based on computational cost (FLOPs) rather than query count. Study of 10 models with 3 attack strategies reveals: alignment has non-monotonic effects, scaling reduces gradient attacks but not template attacks, transfer possible across models, cost varies up to 5× across harm categories.

Summary generated by Claude — human-verified