Researchers Bypass AI Limitations by Overloading It with Complex Jargon

A recent study by researchers from Intel, Boise State University, and the University of Illinois reveals a method to “jailbreak” AI chatbots like ChatGPT and Gemini using a technique called “InfoFlood.” By complicating queries with academic jargon and fictitious references, users can bypass built-in safeguards against harmful or malicious content. The study, titled “InfoFlood: Jailbreaking Large Language Models with Information Overload,” details an automated attack system that rephrases prohibited questions to induce chatbot compliance. This method highlights the susceptibility of large language models (LLMs) to linguistic manipulations, prompting calls for enhanced security measures to resist such adversarial tactics. Researchers claim a high success rate in circumventing existing guardrails by transforming malicious requests into complex, theoretically-focused inquiries. They emphasize the need for stronger defenses against these manipulations to ensure the integrity of AI interactions. The findings suggest a pressing necessity for improved model training and security protocols in AI systems to mitigate these risks effectively.

Source link