OpenAI’s August fix for theft of reasoning traces was quickly bypassed

Research firm Xint has published details of a bug in the OpenAI API that leaks raw reasoning traces from frontier OpenAI models, bypassing mitigations OpenAI deployed in August after an earlier report dubbed ‘Stolen Thoughts’ revealed a similar flaw, CSO Online reports. OpenAI confirmed the flaw, has mitigated it and paid a bug bounty; according to the report, leaked reasoning traces could enable large-scale data extraction and invisible prompt injection. A SecurityWeek version of the same story (idx 138) was dropped as a duplicate version.