OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections

Summary: Security firm Xint showed that OpenAI's compaction feature, which shrinks a long conversation into a smaller encrypted item, could be instructed to copy a model's hidden reasoning into that item in plain text, and the technique got past protections OpenAI added in August against reasoning theft. Xint ran it against gpt-5.6-sol using math and algorithm puzzles its researchers wrote, passed the compacted item to the cheaper gpt-4.1 and gpt-4o models to repeat it, and OpenAI confirmed and fixed the issue on October 23, paying a $6,000 reward.
What's compelling: OpenAI keeps raw reasoning hidden partly to stop distillation, and it said in September it had disrupted a July campaign to expose that reasoning with a core part linked to people associated with China's Moonshot AI - so a working path to copy reasoning out in plain text is a direct route to the training data a rival would want. Xint's broader point is that any server-side step reading decrypted reasoning and following user instructions could leak it, and it did not test whether GPT-6 Astra is affected.

Read the original at The Hacker News ↗