Overview
Severity: HIGH | Affected: Major LLM Providers | Category: research
A paper published by researchers at the Stanford AI Lab details a novel jailbreak technique named 'Contextual Hijacking.' The attack circumvents the safety alignment of major large language models by embedding subtle, adversarial instructions deep within a long and seemingly benign context window. Unlike direct prompt injection, this method gradually shifts the model's behavior over thousands of tokens, causing it to eventually comply with harmful requests without triggering standard safety filters. The paper demonstrates a success rate of over 85% against several popular closed-source models. The technique raises significant concerns for AI applications that rely on processing large documents or maintaining long conversation histories, as the lengthy context becomes a viable attack surface. Model providers are now scrambling to develop defenses against this stealthy and effective new threat vector.