Researchers identify vulnerability in OpenAI model reasoning traces - MSSP Alert

OpenAI has addressed a security vulnerability in its application programming interface that permitted the extraction of raw reasoning traces from frontier models, MSSP Alert reported. The outlet based its account on information published by Infosecurity Magazine. According to the report, a firm called Xint carried out the extraction. The vulnerability described was in the API itself.

The technique used, per the report, is called compaction smuggling. The article states that it bypassed protections intended to hide the chain of thought behind model responses. It also states that OpenAI's August mitigations did not account for compaction processes. The report does not specify which models were affected, how many traces were extracted, when the technique was used, or whether Xint disclosed the issue to OpenAI before or after the mitigations were applied.

Reasoning traces, in the article's framing, are the chain of thought behind model responses — material the protections were designed to conceal. The article presents the episode as a mitigation that was applied and then found not to cover a particular process. It states that OpenAI has addressed the vulnerability, without describing the fix. MSSP Alert's account relies on Infosecurity Magazine's reporting for the details of the technique and for the August timeline. No statement from OpenAI beyond the reported mitigation appears in the excerpt.