OpenAI Fixed Flaw That Exposed Hidden AI Reasoning

Researchers at Xint found a flaw in OpenAI’s API that let them recover reasoning traces the company normally keeps hidden from API users.

The technique, called compaction smuggling, involved hiding one model’s reasoning in a compressed conversation summary and asking another model to reproduce it.

Xint said it recovered reasoning from GPT-5.6 Sol using older GPT-4.1 and GPT-4o models, and the method worked at every reasoning-effort setting it tested.

OpenAI had introduced earlier safeguards in August after researchers showed that hidden reasoning could be replayed into a weaker model, but those safeguards did not cover this use of context compaction.

OpenAI fixed the flaw on October 23 and paid Xint $6,000 for reporting it.

The concern is that stolen reasoning could provide richer material for training rival models than ordinary answers, a process known as distillation.

Xint warned the technique could enable distillation at scale, but did not say it had used the extracted reasoning to train a model.