Executive Summary
Researchers demonstrated that OpenAI's compaction feature could expose hidden reasoning that is normally returned only in encrypted form. Organizations using reasoning models or AI applications should assess whether similar state-management features can leak protected model data.
What Happened
OpenAI's compaction feature, which shrinks a long conversation to save space, could be turned into a way to copy a model's hidden reasoning out in plain text, security firm Xint has shown. OpenAI's reasoning models work through a problem before answering, and the service normally returns only the raw reasoning as an encrypted item the user cannot read. The method got past
Affected Products
- OpenAI compaction feature
- OpenAI reasoning models
Vulnerability / Attack Technique
Hidden reasoning extraction via compaction feature abuse
Technical Mechanism
The compaction feature shrinks a long conversation to save space. According to the summary, researchers used this feature to copy the model's hidden reasoning into plain text, bypassing the normal encrypted representation that users cannot read.
AI Security Category
LLM07:2025 System Prompt Leakage
Relevant Frameworks
RealGround Analysis
This issue matters because conversation compaction and other context-management functions can create new paths for protected model content to become observable. Teams running AI agents or LLM applications should include state transformation, summarization, and encrypted-output handling in adversarial security testing.
Recommended Controls
- Red-team compaction, summarization, and context-window reduction features for hidden-prompt and hidden-reasoning disclosure.
- Enforce server-side isolation of hidden reasoning and prevent transformed conversation state from exposing protected content.
- Add automated output inspection to detect hidden reasoning or system-level content in user-visible responses.
- Review encryption and access-control boundaries for reasoning traces, intermediate states, and conversation artifacts.
- Include model-output confidentiality tests in AI security readiness assessments and regression testing.
Related RealGround Service
AI Security Readiness Assessment · or start with the free AI Agent Security Scanner.
Sources
thehackernews.com: https://thehackernews.com/2026/10/openai-compaction-flaw-let-researchers.html
Summary and analysis written by RealGround from the source above; last updated 2026-10-27.