OpenAI Fixes Compaction Smuggling Flaw That Exposed Hidden AI Reasoning

OpenAI fixed an API vulnerability that allowed researchers at Xint to extract hidden reasoning traces through a technique called Compaction Smuggling. Discovered in August 2026, the flaw let researchers place encrypted reasoning into a conversation compaction and have another model render its contents, bypassing restrictions introduced after earlier research demonstrated cross-model reasoning replay. Xint reported that OpenAI confirmed the issue, deployed an undisclosed fix on October 23, and awarded a $6,000 bounty. Researchers obtained consistent outputs from two decoding models but could not establish token-for-token fidelity to the original reasoning.

The exposure could enable large-scale model distillation, allowing competitors to use proprietary reasoning traces to train cheaper rival systems, Xint warned. SecurityWeek separately cited a CISA warning alleging extensive extraction from U.S. frontier AI models by Chinese AI companies, but neither report established that those companies used Compaction Smuggling. The findings highlight context compaction and cross-model data transfers as security boundaries that AI providers need to protect against proprietary-data extraction.