Vulnerabilities of Language Models: an analysis of ChatGPT

Large language models (LLMs) like ChatGPT have revolutionized human-machine interaction, offering intelligent and relevant responses to a wide range of questions. However, like any advanced technology, these models also present potential vulnerabilities.

This article examines the possible compromises of such systems and the security measures implemented to protect them, highlighting why hacking attacks on these models are extremely unlikely.

Inferential attacks represent one of the most sophisticated threats to LLMs. These attacks aim to extract sensitive information through the aggregation of seemingly innocuous responses. The techniques used include dual-intent questions, designed to indirectly extract sensitive information, and the decomposition of complex questions into simpler sub-questions.

The aggregation of responses to these sub-questions can potentially reveal sensitive details that the model would not provide in response to a direct question.

To counter these threats, ChatGPT implements several advanced security measures. Information censorship limits the loss of sensitive data in the model’s responses, introducing noise or slight modifications to hinder the inference of confidential information. Fine-tuning alignment trains models with human feedback to…