Overview
Severity: HIGH | Affected: OpenAI, Google, Anthropic | Category: research
Researchers at the Stanford AI Lab have published a paper detailing a novel jailbreak technique called 'Semantic Splicing.' This method bypasses safety filters in leading Large Language Models by embedding harmful instructions within complex, semantically-rich narratives. Unlike simple prompt injection, Semantic Splicing crafts a benign-looking story where the harmful request is distributed across multiple, contextually-linked sentences. The model, focused on maintaining narrative coherence, fails to recognize the malicious intent until it has already generated the unsafe output. The paper demonstrates successful attacks against models from Anthropic, Google, and OpenAI, achieving over an 85% success rate in bypassing safety alignments. The researchers have privately disclosed their findings to the affected companies, who are now working to patch their models against this sophisticated, context-aware attack vector. The findings highlight the growing challenge of securing LLMs against attacks that exploit the models' core reasoning capabilities.