OpenAI Patches API Flaw That Let Researchers Extract Encrypted AI Reasoning as Plaintext

OpenAI has patched a critical API vulnerability that allowed researchers to extract encrypted hidden reasoning traces from its frontier models by exploiting the platform's conversation compaction feature, potentially enabling large-scale model distillation. The technique, dubbed "compaction smuggling" by researchers at Xint, was discovered in August 2026 and disclosed publicly on October 26 after OpenAI deployed a fix on October 23.