OpenAI Compaction Feature Exposed Hidden Reasoning in ChatGPT

A security research team from Xint has identified a vulnerability in OpenAI's ChatGPT related to the conversation compaction feature. This feature is designed to compress lengthy conversations to reduce storage and bandwidth costs. However, the researchers found that by manipulating the compaction process, they could bypass the encryption that normally protects the model's hidden reasoning outputs.

OpenAI's reasoning models typically perform a chain-of-thought process before generating a final answer. To protect user privacy, the intermediate reasoning steps are returned as encrypted items that users cannot directly read. The Xint team discovered a method to exploit the compaction feature to retrieve these reasoning outputs in plain text, effectively bypassing the intended obfuscation.

The findings highlight a privacy and security concern where the internal processing of reasoning models could be exposed through a feature intended for operational efficiency. OpenAI has been notified of the issue, and the company is reportedly investigating the vector to determine appropriate mitigations.