Jailbreak defenses are essential for protecting large language models (LLMs), but they can also introduce secondary costs that weaken model utility. We present a systematic study of these defense trade-offs along three dimensions: performance impact, over-refusal on benign inputs, and inference cost. Rather than treating defenses as a single class, we organize them by operational strategy and examine how different strategies correlate with different side-effect profiles. Across state-of-the-art defense methods, widely used benchmark datasets, and representative open-source LLMs, we find that defenses rarely improve downstream capability, but instead vary in how they trade safety gains against usability and efficiency. In particular, rule-based defenses best preserve task performance, highly conservative self-reflective defenses often increase over-refusal, and multi-round defenses incur the largest runtime overhead. These results provide both a benchmark for evaluating defense side effects and practical guidance for selecting defenses under deployment constraints.
When LLM Defenses Backfire: Characterizing Safety, Performance, and Cost Trade-offs
Jailbreak defenses are essential for protecting large language models (LLMs), but they can also introduce secondary costs that weaken model utility. We present a systematic study of these defense trade-offs along three dimensions: performance impact, over-refusal on benign…
- Preview

- Year
- 2026
- Hosting
- Full text hostedCC-BY-4.0
Cite
Notes
Only stored in your browser.
Attribution
- Abstract & full text
- arxiv.org/abs/2607.24392CC-BY-4.0
- TL;DR
- Semantic Scholar