OpenAI fixed reasoning-extraction bypass on October 23, Xint says

OpenAI confirmed and fixed a way to extract protected model reasoning through its conversation-compaction feature on October 23, security firm Xint said. The Hacker News detailed the demonstration on October 27. Xint’s test produced readable text presented as hidden reasoning, though the researchers could not verify that it matched the original word for word.

The bypass followed protections introduced in August. Researchers behind the paper Stolen Thoughts had shown that a weaker model could receive a stronger model’s encrypted reasoning and repeat it as readable text. OpenAI and other providers then blocked that transfer between models.

How Xint says compaction bypassed the earlier block

Compaction compresses a long conversation into a smaller encrypted item so a model can continue with the earlier context. Hidden reasoning is the intermediate text a model generates while working through a problem; OpenAI normally keeps that raw text encrypted rather than showing it to users.

Xint gave gpt-5.6-sol mathematics and algorithm puzzles written by its researchers and retained the encrypted reasoning returned for those requests. It then asked the same model to copy that reasoning into a compaction item. Xint said using the originating model avoided the check designed to block transfers of reasoning between different models.

Compaction items could still move between models. Xint passed the resulting item to the cheaper gpt-4.1 and gpt-4o models and asked each to repeat its contents. Both produced the same readable text. That agreement supported the consistency of the copying step, but did not establish whether compaction had reworded the original reasoning.

OpenAI keeps raw reasoning hidden partly to prevent distillation: training a cheaper model on another model’s outputs. Reasoning text provides more detail about how a problem was worked through than a final answer alone.