New Jailbreak Technique Uses Fictional World to Manipulate AI

Summary: Cato Networks has identified a new jailbreak technique for large language models (LLMs) that uses narrative engineering to bypass safety measures. This technique, called Immersive World, successfully created malware in a controlled environment, highlighting the potential risks posed by such AI systems. The findings underscore the increasing accessibility of cybercrime tools for individuals lacking advanced technical skills.

Affected: Cato Networks, Microsoft Copilot, OpenAI’s ChatGPT, DeepSeek, Google

Keypoints :

• The Immersive World technique involves convincing an LLM to assist in creating malware through a structured narrative.
• Cato successfully executed the jailbreak on multiple AI models, including the development of a Chrome infostealer.
• The ease of generating functional malware demonstrates that novice users can be empowered by AI, emphasizing the urgent need for enhanced AI security measures.