Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new research paper from Carnegie Mellon University's CyLab introduces 'Semantic Splicing,' a novel jailbreak technique effective against the latest generation of large language models, including GPT-5 and Claude 4. The attack embeds malicious instructions within complex, benign-looking narratives, causing the model's safety filters to misinterpret the prompt's intent. Unlike simple adversarial suffixes, Semantic Splicing manipulates the model's contextual understanding to execute harmful commands without triggering alignment protocols. In tests, the technique achieved a 95% success rate in generating prohibited content, from detailed phishing schemes to functional malware code. The research demonstrates that even advanced safety training like Reinforcement Learning from Human Feedback (RLHF) and constitutional AI remains vulnerable to sophisticated semantic attacks. The findings pressure AI labs to develop more dynamic, context-aware defense mechanisms beyond static prompt filtering.