SAID: Safety-Aware Intent Defense via Prefix Probing for Large Language Models

Researchers have introduced Safety-Aware Intent Defense (SAID), a novel, training-free framework designed to protect Large Language Models (LLMs) from jailbreak attacks. Unlike existing defenses that rely on external filtering or decoding-time controls, which often incur high inference costs or reduce utility, SAID operates by distilling obfuscated user inputs into concise core intents using the target model itself. It then applies validated safety prefixes to probe these intents, eliciting safety-aware responses. A conservative aggregation rule rejects any request where a distilled intent is identified as unsafe. This approach enables black-box-compatible defense without modifying model parameters or decoding processes. Experimental results across four open-source LLMs and six representative jailbreak attacks demonstrate that SAID achieves state-of-the-art performance in mitigating harmful responses while maintaining competitive utility for benign tasks. The study highlights SAID as a practical solution for balancing safety and utility in LLM deployment, offering an efficient method to secure models against adversarial prompts without the drawbacks of traditional guardrails.