Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A paper published by researchers at the Stanford Artificial Intelligence Laboratory (SAIL) introduces a novel jailbreak technique named 'Cognitive Juggling'. The attack bypasses the safety alignment of major large language models by crafting prompts that force the model to simultaneously track multiple, contradictory, and complex conceptual states. This induced cognitive load causes a cascading failure in the model's safety filters, allowing it to respond to malicious instructions that would normally be refused. Unlike previous methods that rely on long context windows or role-playing, Cognitive Juggling is highly efficient and works with short, concise prompts, making it difficult to mitigate with simple input filtering. The research team responsibly disclosed the vulnerability to leading AI labs, including OpenAI, Google, and Anthropic, 90 days prior to publication, and all have acknowledged the issue, pushing emergency patches to their models.