Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A paper published by researchers at Carnegie Mellon University introduces 'Semantic Splicing,' a new jailbreak technique highly effective against the latest generation of large language models. Unlike traditional prompt injection that relies on confusing the model with conflicting instructions, Semantic Splicing embeds harmful requests within complex, benign-sounding narratives or role-playing scenarios. The technique works by building a deep semantic context that makes the malicious instruction appear as a logical continuation of the safe narrative, thereby bypassing the model's alignment training and safety guardrails. The researchers demonstrated a 92% success rate in generating harmful content across models from OpenAI, Google, and Anthropic. The findings highlight a fundamental weakness in current content filtering mechanisms, which primarily focus on explicit keywords and simple instructions rather than nuanced contextual deception.