Overview
Severity: HIGH | Affected: All Major LLMs | Category: research
A new research paper from the Stanford AI Lab details a powerful jailbreak technique named 'Cognitive Jigsaw.' The attack works by splitting a malicious request into multiple, seemingly benign and logically disconnected prompts. These individual prompts are processed by the LLM in separate contexts. A final, cleverly crafted prompt then instructs the model to synthesize the outputs from the previous turns, effectively reassembling the 'jigsaw puzzle' into a coherent, harmful response that bypasses the model's safety filters. The technique has proven effective against leading models like GPT-5, Claude 4, and Gemini Ultra, achieving a success rate of over 85% in red-teaming trials. This research exposes fundamental vulnerabilities in context-based safety mechanisms and forces a re-evaluation of how models handle multi-turn conversational memory.