Overview
Severity: HIGH | Affected: Multiple LLMs | Category: research
Researchers at Carnegie Mellon University published a paper detailing a novel jailbreak technique named 'Semantic Splicing.' This method circumvents safety filters on leading large language models, including GPT-5 and Claude 4, by embedding harmful instructions within seemingly benign, complex narratives. The technique splices malicious commands across different semantic contexts, making them difficult for alignment models to detect as a coherent harmful request. For example, a request to generate malware code is disguised as a creative writing prompt about a fictional cybersecurity expert. The researchers demonstrated a success rate of over 85% against leading commercial models, raising significant concerns about the robustness of current safety alignment methods and prompting calls for more advanced, context-aware defense mechanisms.