OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections

Researchers at security firm Xint discovered a flaw in OpenAI's compaction feature that allows them to bypass protections and access a model's hidden reasoning in plain text. Normally, OpenAI's reasoning models work through a problem before answering and only return the raw reasoning as an encrypted item that users cannot read. The compaction feature, which is intended to shrink long conversations to save space, can be exploited to copy this hidden reasoning. This method was able to get past the usual protections that prevent users from accessing the raw reasoning. The vulnerability allows the hidden reasoning to be output in plain text, rather than being encrypted and inaccessible to users.