AI Runtime Security Tools

Study flips Prompt Guard 2 verdicts with paraphrases

An arXiv case study posted 21 Sep 2026 applied explainability methods to Meta's Prompt Guard 2 classifier. It found decisions rest on many tokens, yet synonym substitution and paraphrasing could flip verdicts, and it argues that transparency aids can also make bypasses cheaper to build. The paper proposes no new product.

Facts read on .

On 21 Sep 2026, researchers posted "Decoding Guardrails: XAI-Guided Perturbation Analysis of Prompt Injection Detection" to arXiv. It is an exploratory case study of how Prompt Guard 2, a classifier-based guardrail widely used as a first line of defense, separates malicious from benign prompts.

Using Vanilla Gradient and SHAP attributions, the authors find the classifier's decisions rely on the combined contribution of many tokens rather than a few. Even so, saliency-guided synonym substitution and sentence-level paraphrasing can flip its predictions while changing only a moderate fraction of the text, and in some cases the result was a successful jailbreak against the underlying model. A dataset-scale analysis shows that undetected injection prompts systematically lack the lexical markers the classifier relies on.

The paper's design conclusion is for builders of classifier-based guardrails: evaluate against paraphrase and substitution, and recognize that explanation methods meant to support transparency can lower the cost of building bypasses. For buyers, it supports gating the action an injection asks for as well as scoring the text.

Sources