AnswersPublished17 sources
How Compaction Smuggling Exposed OpenAI’s Hidden Reasoning
Xint reported that compaction smuggling let researchers route hidden reasoning from an OpenAI model through a compaction item and have another model transcribe it. The issue highlights a security gap in how protected model data is handled across API features—not evidence that encryption itself was broken.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Xint reported that compaction smuggling let researchers route hidden reasoning from an OpenAI model through a compaction item and have another model transcribe it.
- The issue highlights a security gap in how protected model data is handled across API features—not evidence that encryption itself was broken.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How did security firm Xint’s three-call “compaction smuggling” attack extract hidden reasoning from OpenAI’s gpt-5.6-sol as plain text—by re. Article summary: Xint reported that “compaction smuggling” turned OpenAI’s conversation-compaction feature into a route for exposing hidden reasoning as plain text, and Tim Becker claimed it bypassed all mitigations on the frontier model. Topic tags: general, general web, user generated, government. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, ch
Xint reported that a flaw in OpenAI’s conversation-compaction workflow let researchers expose hidden reasoning from a model as readable text. The technique, called compaction smuggling, reportedly placed reasoning in a compaction item that another model could read back. OpenAI fixed the issue on October 23, 2026, according to the company’s disclosure and news reports. 2
5
How compaction smuggling worked
Conversation compaction reduces a long chat to a shorter representation. Xint’s report describes steering that process to copy hidden reasoning into a compaction item, then passing the item to an older model that could transcribe its contents. 5
43
That matters because the reasoning was not simply made visible through a normal model response. Reporting describes reasoning as normally returned in an encrypted item that users cannot read; the reported flaw used another stage of API processing to make its contents readable. 2 In other words, the concern was how protected data was handled across features—not a demonstrated break of the encryption itself.
Some secondary reporting identifies gpt-5.6-sol as the model whose reasoning was extracted and says the compaction output was echoed through gpt-4.1 and gpt-4o. 33
34 The publicly available source excerpts do not substantiate every detail of the purported three-call sequence, including the exact formatting or classifier-evasion technique, so those specifics should be treated cautiously.
Why exposed reasoning could matter
Xint said the vulnerability could enable large-scale distillation of advanced models. 5 In distillation, outputs from a stronger model can be used as training material for a smaller one. Reasoning traces could potentially provide more information about how a model arrived at an answer than the answer alone, making them valuable training targets. That describes a potential use of the exposed data; the reporting does not establish that anyone used this flaw to train a competing model.
What is known about the fix—and what remains unclear
Xint said OpenAI confirmed the issue and deployed an undisclosed fix on October 23, before Xint published its report on October 26. 5 The Hacker News also reported that OpenAI confirmed and fixed the problem.
2 A secondary report put the bug bounty at $6,000.
50
The available reporting does not describe the fix in enough detail to assess how it changed the API’s handling of compaction items. Nor does it establish the precise attack formatting or verify claims about testing across every reasoning-effort level. Those limits matter: the finding supports concern about this reported route, not a conclusion that every OpenAI model or workflow was affected in the same way.
The security lesson: retest across the whole workflow
The incident illustrates why protecting a sensitive output at one API boundary may not be enough if another feature can process or transfer it. Xint said the technique bypassed mitigations OpenAI had introduced after earlier research on reasoning-trace theft. 5 CSO Online highlighted the broader security lesson: a fix should be retested because it may not fully resolve a model-related problem.
3
For AI providers, the key question is not only whether a user can directly request hidden reasoning. It is also whether that data remains protected as it passes through features such as compaction and between models. The public information confirms the reported flaw and a vendor fix, but leaves important implementation details undisclosed.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "How Compaction Smuggling Exposed OpenAI’s Hidden Reasoning"?
Xint reported that compaction smuggling let researchers route hidden reasoning from an OpenAI model through a compaction item and have another model transcribe it.
What are the key points to validate first?
Xint reported that compaction smuggling let researchers route hidden reasoning from an OpenAI model through a compaction item and have another model transcribe it. The issue highlights a security gap in how protected model data is handled across API features—not evidence that encryption itself was broken.
Sources
thehackernews.comOpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections
xint.ioStealing OpenAI Reasoning Traces via Compaction ...
github.comGitHub - theori-io/ghost-compact: OpenAI compaction research: GhostCompact + Compaction Smuggling