Xint finds OpenAI flaw exposed hidden AI reasoning traces

Researchers at AI vulnerability firm Xint found that a flaw in OpenAI’s API allowed them to recover raw reasoning traces that the company normally withholds from API users. Xint disclosed the technique, called “compaction smuggling”, on 26 October 2026. It said OpenAI confirmed the issue, fixed it on 23 October and paid a $6,000 bug bounty; the fix has not been publicly described.

The flaw involved OpenAI’s context compaction feature, which uses a model to compress a long conversation into an encrypted item that users cannot read. Xint said it generated reasoning with GPT-5.6 Sol, compacted it, then asked the older GPT-4.1 and GPT-4o models to repeat their input word for word, recovering the hidden reasoning. The researchers said the method worked at every reasoning-effort setting they tested.

OpenAI had added mitigations in August after research showed encrypted reasoning could be replayed into less capable models, but Xint said those measures did not cover compaction on the model that produced the reasoning.

Xint warned that the technique could enable large-scale distillation of OpenAI’s frontier models, but did not say it had trained a model using the extracted material. The researchers cited earlier reports of alleged distillation campaigns and OpenAI’s disclosure that it had disrupted an effort to extract hidden reasoning.