Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new paper from the Stanford AI Lab has introduced a sophisticated jailbreak technique named 'Cognitive Jigsaw.' The attack bypasses the safety filters of leading Large Language Models (LLMs) by distributing a malicious instruction across multiple, seemingly benign prompts. The model's large context window inadvertently reassembles these fragments into a coherent, harmful command, executing it without triggering standard safety protocols. The researchers demonstrated the technique's effectiveness against several major closed-source models, successfully generating disallowed content related to misinformation, hate speech, and weapon creation. The findings challenge current defense mechanisms that focus on analyzing individual prompts, revealing a critical vulnerability in how models process and integrate contextual information. The lab has privately disclosed the full details to affected vendors to allow them to develop mitigations.