Overview
Severity: HIGH | Affected: Multiple LLMs | Category: research
A paper published by researchers at Carnegie Mellon University has introduced a novel jailbreak technique named 'Cognitive Jigsaw'. This method effectively bypasses the safety alignment of state-of-the-art AI systems by distributing a malicious request across multiple, seemingly independent AI agents. The attack works by breaking down a harmful prompt into several benign-looking sub-tasks. Each sub-task is given to a separate agent, which processes its piece without any context of the overall malicious goal. A final orchestrator agent is then prompted to 'assemble the puzzle' from the outputs, leading to the generation of harmful content that would have been blocked if submitted as a single query. The research demonstrates a fundamental vulnerability in systems that rely on per-prompt safety checks, especially in emerging multi-agent and chain-of-thought architectures. The findings challenge current AI safety paradigms and suggest a need for more holistic, context-aware defense mechanisms.