OpenAI Compaction Flaw Exposes Hidden Reasoning Protections

Researchers at Xint extracted OpenAI's encrypted hidden reasoning content as plain text by exploiting a server-side compaction feature, and OpenAI paid the team a $6,000 bounty for the disclosure. The finding exposes a vulnerability in how OpenAI's system handles protected reasoning data.

Researchers at Xint discovered a clever way to tap into OpenAI's hidden reasoning processes, exploiting a server-side compaction feature to extract encrypted model insights and turning them into plain text. They were rewarded with a $6,000 payout for their trouble, a nice payday for exposing a vulnerability in OpenAI's system.