OpenAI's Compaction Flaw Lets Researchers Extract Hidden Reasoning
Original source
OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections
The Hacker News →Security firm Xint demonstrated how to extract hidden reasoning from OpenAI’s reasoning models by uncovering a flaw in the company’s compaction feature. OpenAI’s reasoning models work through problems internally, returning reasoning as encrypted data users normally cannot read. Xint showed how the compaction process—designed to shrink conversations for space efficiency—could be manipulated to convert this encrypted reasoning into readable text.
Xint extracted encrypted reasoning from gpt-5.6-sol and instructed the compaction feature to copy that reasoning into a compacted item while replacing spaces with tabs to bypass OpenAI’s content filters. Because compaction items can move between models, Xint passed these to gpt-4.1 and gpt-4o, which repeated the text as instructed. This circumvented OpenAI’s August protections blocking cross-model reasoning replay, which followed the “Stolen Thoughts” research. Xint acknowledges it cannot verify the extracted text matches the model’s original reasoning exactly, and it did not attempt to access other users’ reasoning. It remains unknown whether anyone used the method before OpenAI patched it.
OpenAI fixed the flaw on October 23 and awarded Xint a $6,000 bounty. The company encrypts reasoning partly to prevent distillation, where competitors train cheaper models on another company’s outputs. Xint tested only gpt-5.6-sol and did not check whether GPT-6 Astra is affected. Though this was a controlled demonstration rather than a real-world attack, the researchers argue any server-side process reading decrypted reasoning and following user instructions could potentially leak it. China-based labs have separately targeted American models’ hidden reasoning, according to reports from Anthropic and a U.S. government advisory, though Xint’s method is unrelated to those efforts.
Read the full article
Continue reading at The Hacker News →This is an AI-generated summary. Read the original for the full story.