Guardrail Models Collapse under Adversarial Paraphrases
Abstract
AbstractWe study the vulnerability of guardrail models to adversarial paraphrases that exploit superficial textual variation: meaning-preserving changes that alter expressions spuriously correlated with safety labels. While prior work probes this vulnerability through simple heuristics, such as keyword-label co-occurrence statistics, we argue that guardrails may rely on more complex textual expressions involving phrases or syntactic structures that cannot be captured by heuristics. To enable systematic evaluation of such hidden vulnerabilities, we propose ASPro, an agentic adversarial evaluation framework that automatically synthesizes meaning-preserving paraphrases containing expressions that contradict guardrails’ learned spurious expression-label correlations. Applying ASPro to five safety benchmarks and 13 guardrail models, we find that the average F1 drops from 77.83 to 40.25, substantially more than under existing heuristic baselines. Moreover, on ASPro prompts, guardrail models often fail to outperform their base LLMs before safety fine-tuning. These results suggest that current safety fine-tuning may overfit to spurious expressions rather than truly improve guardrails’ understanding of prompt safety.