Researchers extract hidden GPT-5.6 reasoning through an OpenAI API flaw
Original title: Researchers Exploit OpenAI API Flaw to Extract Hidden GPT-5.6 Reasoning Traces
OpenAI confirmed the bug, paid Xint a $6,000 bounty, and deployed an undisclosed fix on October 23.
Xint says OpenAI's conversation-compaction feature could read decrypted GPT-5.6 reasoning and copy it into an encrypted item that older models then echoed as text. The method, compaction smuggling, bypassed controls added in August 2026. OpenAI confirmed the bug and fixed it on October 23. Xint used its own requests and warns the traces could aid unauthorized distillation.
Unconfirmed: Xint could not verify that the echoed text exactly matched the original hidden reasoning.
- Xint found the issue in August and published its report on October 26.
- The method worked at every reasoning effort level the team tested.
- One sample had 405 reasoning tokens and 609 echoed output tokens.
- Xint could not prove the output exactly matched the hidden reasoning.
- Vendors
- OpenAI
- Products
- OpenAI API
- Organizations
- Xint
- AI models
- GPT-5.6gpt-5.6-solgpt-4.1gpt-4o
From the original article
2 min read at cybersecuritynews.com
Researchers at Xint found an OpenAI API flaw that exposed hidden reasoning in GPT-5. 6 models by routing prompts through a conversation-compaction feature. The method, called compaction smuggling, bypassed protections introduced in August 2026.
By Abinaya · cybersecuritynews.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.