Overview
Severity: HIGH | Affected: Major LLM Providers | Category: research
Researchers from a leading academic AI lab have published a paper detailing a novel jailbreak technique named 'Semantic Splicing.' The method circumvents safety filters on major large language models by embedding malicious instructions within seemingly benign, but semantically-linked, chunks of text spread across a long context window. The model's attention mechanism inadvertently reassembles the harmful command during processing, bypassing conventional prompt filtering and refusal training. The paper demonstrates successful exploits against several top-tier models, generating harmful content and executing unauthorized code-like instructions. The technique highlights a fundamental vulnerability in how models process distributed context, forcing providers like OpenAI and Google to re-evaluate their architectural safeguards. The research team has responsibly disclosed the vulnerability to affected companies ahead of publication, and patches are reportedly in development, though the core issue may require significant model retraining to fully mitigate.