Skip to content
LowNewsLLM-specific

Stolen AI Reasoning Could Help China Build Cheaper Rivals

Published
Record updated
View JSON

Summary

Xint reports a flaw in OpenAI's context compaction process that lets researchers move reasoning traces from one model to another. The technique, which Xint calls Compaction Smuggling, works by sending a compaction from one model to a different model to output its input as-is. OpenAI confirmed the issue, fixed it, and awarded Xint a $6,000 bounty on October 23, 2026.

Mitigation

OpenAI confirmed the new issue of compaction smuggling, fixed it and awarded Xint a $6,000 bounty on October 23, 2026.