Researchers from Xint have identified a vulnerability in the OpenAI API, specifically within the GPT-5.6 models, that inadvertently revealed hidden reasoning processes through a feature designed for conversation compaction.
Discovery of the Vulnerability
The technique, termed compaction smuggling, effectively circumvented security measures that were implemented by OpenAI in August 2026. The flaw was confirmed by OpenAI, who implemented an undisclosed fix on October 23 and compensated the research team with a $6,000 bounty for their discovery.
This incident highlights how features intended to maintain conversation context can inadvertently bypass security controls. Xint discovered this issue in August and released a detailed technical report on October 26.
Technical Insights into the Exploit
OpenAI’s reasoning models are designed to process problems internally before providing answers, with summaries available to API users while the detailed reasoning remains encrypted. Xint found that through specific API requests, decrypted reasoning could be accessed.
The compaction feature, which condenses lengthy conversations into encrypted items for context preservation, was found to be exploitable. By following specific instructions, users could read decrypted reasoning, creating a potential security breach.
Implications and Response
The researchers demonstrated the vulnerability by sending math and algorithm puzzles to the gpt-5.6-sol model and manipulating API settings to retain encrypted content while bypassing certain security filters. This was then processed by older models, exposing the reasoning as readable text.
Despite protections introduced after previous research on hidden reasoning leaks, this method showed that reasoning traces could still be extracted and potentially used for unauthorized training of other models.
While Xint’s findings did not involve unauthorized model use, they underscore the need for robust security measures for server-side features handling decrypted reasoning, even when outputs are encrypted.
As artificial intelligence continues to evolve, so do the potential for cyberattacks, necessitating vigilant defenses capable of responding at machine speed.
