Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new paper published by researchers at the Stanford AI Lab details a novel jailbreak technique named 'Cognitive Override.' Unlike traditional prompt injection attacks that rely on single, cleverly crafted inputs, this method uses a multi-turn conversational approach to construct a complex, synthetic narrative. By establishing a fictional context where a harmful request is framed as a benign or necessary action (e.g., a safety test scenario or a fictional story plot), the technique effectively bypasses the safety alignment of major large language models. The research demonstrated a success rate of over 85% against leading models, including GPT-5 and Claude 4, raising significant concerns about the robustness of current content filtering and alignment strategies that focus primarily on analyzing individual prompts rather than the holistic conversational context.