Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new research paper from Carnegie Mellon University's CyLab introduces a novel jailbreak technique named 'Semantic Splicing'. This method circumvents the safety filters of leading Large Language Models by embedding harmful prompts within complex, benign-looking narratives. The technique works by creating a dual-layered semantic structure in the prompt; the model's safety alignment system processes the harmless surface layer, while the underlying 'spliced' semantics trigger the generation of prohibited content. The researchers demonstrated an over 85% success rate against several state-of-the-art models from companies like Google and Anthropic, successfully generating outputs ranging from malicious code to detailed misinformation. This attack vector exposes a fundamental weakness in current alignment strategies that struggle to interpret deeply nested user intent, prompting calls for more robust, context-aware safety mechanisms.