OpenAI API compaction flaw leaked hidden model reasoning traces
Original title: Researchers Discover OpenAI API Vulnerability That Leaks Hidden AI Reasoning Traces
Xint says OpenAI confirmed the bug, paid $6,000, and shipped an undisclosed fix on October 23.
Xint researchers disclosed Compaction Smuggling, a three-call OpenAI API technique that moved protected reasoning from gpt-5.6-sol into encrypted compaction output that gpt-4.1 and gpt-4o reproduced as readable text. The issue, found in August, bypassed limits on replaying encrypted reasoning across models by using server-side compaction rather than breaking encryption. OpenAI confirmed it and deployed a fix on October 23, 2026. A demonstration script is on the project's public GitHub.
Unconfirmed: Researchers could not confirm token-for-token fidelity without the original plaintext.
- The attack worked at every reasoning-effort setting tested, from minimal to maximum.
- A 405-token reasoning item became a 612-token compaction item and 609-token echoes.
- Replacing spaces with tabs helped bypass usage-policy classifiers, Xint reported.
- Earlier mitigations had blocked direct reasoning replay across models.
Impact: Hidden reasoning traces recovered as readable text
| Product | Affected | Fixed in |
|---|---|---|
| OpenAI API reasoning models | gpt-5.6-sol, decoded with gpt-4.1 and gpt-4o | undisclosed fix on October 23, 2026 |
- 1.Rely on OpenAI's October 23 server-side fix; a public demonstration script exists.
- Vendors
- OpenAI
- Products
- OpenAI API
- Organizations
- Xint
- AI models
- gpt-5.6-solgpt-4.1gpt-4o
From the original article
2 min read at gbhackers.com
Researchers have disclosed a vulnerability in the OpenAI API that reveals hidden reasoning traces by exploiting conversation compaction. This technique, named “Compaction Smuggling,” bypassed existing safeguards and could allow for the extraction of valuable training material from advanced models. According to Xint, OpenAI confirmed the issue, implemented an undisclosed fix on October 23, 2026, and […]
By Divya · gbhackers.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.