AI Chatbots at Risk: Information Overload Exploits Revealed

With artificial intelligence (AI) becoming increasingly integrated into daily life, concerns about its security vulnerabilities are growing. A recent collaborative study by researchers from Intel, Boise State University, and the University of Illinois has uncovered a critical weakness in large language models (LLMs): their susceptibility to information overload attacks.

Image source note: The image was generated by AI, and the image licensing service provider is Midjourney

The InfoFlood Attack System

The research team developed an automated attack system named InfoFlood, designed to overwhelm AI chatbots with excessive information. This method exploits the models' inability to process overloaded inputs effectively, causing their safety filters to fail.

Key findings include:

• LLMs may take defensive measures under normal circumstances but become vulnerable when flooded with data.
• The attack uses a standardized prompt template containing task definitions, rules, context, and examples.
• When an AI model refuses a query, InfoFlood repeatedly feeds it more information until it complies.

How the Exploit Works

The technique involves:

1. Rule manipulation: Using false citations and fabricated research to align with malicious prompts.
2. Language transformation: Carefully rewording queries to remove overtly harmful intent while maintaining the underlying goal.
3. Overloading context: Flooding the model with excessive data to confuse its filtering mechanisms.

Despite built-in safeguards in models like ChatGPT and Gemini, the study demonstrates that these protections can be circumvented through systematic information overload.

Implications for AI Security

The research highlights significant challenges for AI developers:

• Current safety filters may not adequately handle complex or overloaded inputs.
• Malicious actors could exploit this vulnerability to generate harmful content.
• The findings suggest that LLMs may not fully comprehend user intent when processing dense information.

The team plans to share their findings with companies deploying LLMs, urging them to strengthen their security protocols against such attacks.

Key Points:

📌 Vulnerability exposed: LLMs can be tricked into answering dangerous questions via information overload. 📌 Automated attacks: The InfoFlood system automates exploitation of this weakness. 📌 Filter failure: Safety measures fail when models are overwhelmed with data.