Re-Triggering Safeguards: Embedding Disruption Method Reactivates Built-In LLM Safety

arXiv·medium signal

This paper proposes a jailbreak detection method that works by disrupting embeddings to re-activate the model's built-in safeguards rather than serving as a standalone defense. The insight is that jailbreaking prompts are inherently fragile — small perturbations in embedding space cause the model to recognize the harmful intent it was tricked into missing. Unlike prior defenses, this approach is complementary to existing safety layers rather than replacing them.