New 'Compaction Smuggling' technique bypassed OpenAI's defenses against stolen model reasoning
Security firm Xint said it found a flaw in OpenAI's compaction process, the step where a conversation's accumulated history is compressed into a summary before it overflows a model's context window, that allowed a model's hidden internal reasoning to be moved into another model, a technique that could significantly cut the cost of building a rival AI through distillation (training a smaller model to mimic a larger one's abilities). The technique, which Xint calls Compaction Smuggling, bypassed defenses OpenAI had already put in place in August after earlier research on stolen reasoning, since those defenses only blocked replaying reasoning into a different model, not within the same model that produced it, which is where the new compaction call runs. OpenAI confirmed the issue, fixed it, and paid Xint a $6,000 bounty on October 23, while US agency CISA had already warned in September that Chinese companies including DeepSeek, Alibaba and Moonshot AI have extracted billions of tokens from Western frontier models to train their own. For a business that builds its product on a proprietary AI model through an API, the case shows that protecting a model's internal reasoning is an ongoing race of fixes, not a one-time solution.
Sources
This summary was written automatically with AI from the public sources listed above and checked by an independent AI reviewer before publication. For full details, read the original source.