Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new research paper from Stanford University's AI Lab details a sophisticated jailbreak technique named 'Contextual Weaving.' The attack bypasses the safety alignment of major large language models (LLMs) like GPT-5 and Claude 4 by embedding harmful instructions within complex, benign-looking narratives and multi-turn conversations. Instead of direct commands, the attack gradually builds a conversational context that frames the malicious request as a logical and harmless continuation of the discussion, effectively tricking the model's safety filters. The researchers demonstrated a 92% success rate in generating harmful content, including misinformation and malicious code, across several leading commercial and open-source models. The paper calls for more robust, context-aware safety mechanisms that can analyze the semantic drift of a conversation over multiple turns, rather than just evaluating individual prompts. The findings have prompted major AI labs to re-evaluate their defense strategies against such nuanced adversarial attacks.