OpenAI compaction flaw let researchers extract hidden model reasoning
Original title: OpenAI Compaction Flaw Exposed AI Reasoning to Researchers
Xint hid GPT-5.6 Sol reasoning in a compaction and recovered it through GPT-4.1 and GPT-4o; OpenAI paid $6000.
Xint disclosed compaction smuggling, a flaw in OpenAI's API context compaction that recovered raw reasoning traces the company normally withholds from API users. Researchers generated reasoning with GPT-5.6 Sol, compacted it, and had GPT-4.1 and GPT-4o repeat their input word for word. Xint said this worked at every reasoning-effort setting tested. OpenAI confirmed the issue, deployed a fix on October 23, and paid a $6000 bounty; the fix was not publicly described.
Unconfirmed: OpenAI's fix was not publicly described, and Infosecurity had not received a company statement.
- Xint found the issue in August and published the research on October 26.
- August mitigations blocked replaying reasoning into other models but not compaction.
- Xint did not report training a model on the extracted reasoning.
- OpenAI said an unreleased model had rarely written instructions into its own compaction summaries.
Impact: Recovery of raw chain-of-thought OpenAI withholds from API users
| Product | Affected |
|---|---|
| OpenAI API context compaction | — |
- Vendors
- OpenAI
- Products
- OpenAI API
- Organizations
- Xint
- AI models
- GPT-5.6 SolGPT-4.1GPT-4o
From the original article
2 min read at infosecurity-magazine.com
Researchers uncovered an OpenAI API flaw that exposed hidden AI reasoning traces, prompting a fix and a $6000 bug bounty reward
By Alessandro Mascellino · infosecurity-magazine.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.