Security firm Xint has demonstrated that OpenAI's compaction feature, designed to reduce the size of extensive dialogues, can be exploited to extract a model's internal reasoning in unencrypted text.
OpenAI's reasoning models process a problem prior to generating a response. Typically, the system only provides the raw reasoning, which is delivered to the user as an encrypted element that is inaccessible to them.
The technique developed by Xint successfully circumvented these existing safeguards.
This vulnerability impacts the operational integrity of the reasoning pipeline within the platform.
The ability to bypass the intended encryption mechanism raises concerns regarding proprietary model outputs.
This development highlights potential exposure points in conversational AI infrastructure where internal processing steps are managed.
