Researchers bypass OpenAI's August fix and steal model reasoning traces

Xint disclosed an OpenAI API flaw, dubbed compaction smuggling, that leaks raw reasoning traces from frontier models. The technique bypasses OpenAI's August mitigations by replacing spaces with tabs and running compaction on the same model. OpenAI confirmed and fixed the bug and paid a bounty.

• Swapping spaces for tabs slipped past OpenAI's usage-policy classifiers
• The attack worked at every reasoning effort level tested
• OpenAI confirmed the flaw, mitigated it and paid a bug bounty
• Experts urge CISOs to demand variant testing from AI vendors