Xint finds Compaction Smuggling attack that steals OpenAI reasoning traces

Researchers at Xint discovered a Compaction Smuggling attack that extracts encrypted reasoning traces from OpenAI frontier models and can be used to cheaply distill rival AI. OpenAI confirmed and fixed the issue, paying a $6,000 bounty on October 23, 2026.

• The attack bypasses OpenAI's August mitigations that blocked replaying reasoning into other models
• The three-step scheme obtains reasoning, shapes the compaction, then sends it to another model
• Xint says the method worked at every reasoning effort it tried
• CISA previously reported billions of tokens extracted by Chinese firms from US frontier models