OpenAI has patched a vulnerability in its compaction feature that was exploited by security firm Xint to reveal hidden reasoning from its GPT-5.6-sol model. The compaction feature is designed to condense lengthy conversations into smaller, encrypted items for efficiency. However, Xint discovered a method to bypass encryption safeguards, allowing them to extract and replicate the model's reasoning in plain text. OpenAI confirmed the issue and implemented a fix on October 23.

The flaw was identified during Xint's tests involving math and algorithm puzzles. While the extracted reasoning appeared consistent with what the model processed, Xint acknowledged they could not verify if the text was an exact match to the original reasoning. The technique worked by employing compaction to convert encrypted reasoning into a compacted item, which was then repeated by cheaper models like GPT-4.1 and GPT-4o.

This vulnerability bypassed a prior fix aimed at preventing cross-model reasoning replay, as the compaction process occurs within the same model. Xint's method expanded on their earlier GhostCompact findings, demonstrating the potential to manipulate the compaction step for unauthorized tasks. Although OpenAI swiftly addressed the vulnerability, the incident underscores the risks associated with exposing raw reasoning, which could be exploited for distilling competitive models.

The issue was limited to the GPT-5.6-sol model, and it remains unclear if newer models like GPT-6 Astra were affected. OpenAI's prompt response and undisclosed fix ensured developers using the API need not make changes. Xint received a $6,000 reward for reporting the vulnerability, emphasizing the importance of collaborative efforts in enhancing AI security.