OpenAI patches compaction smuggling used to steal model reasoning
Original title: Stolen AI Reasoning Could Help China Build Cheaper Rivals
Xint received a $6,000 bounty on October 23 after OpenAI confirmed and fixed the bypass of its August mitigations.
Xint says Compaction Smuggling can move one OpenAI model's reasoning into another by encoding it inside a compaction summary, bypassing August controls that blocked direct replay. The firm showed a path from gpt-5.6-sol into gpt-4.1 and gpt-4o and said it enables large-scale distillation. OpenAI confirmed the issue and fixed it. CISA has separately warned that Chinese AI firms extracted billions of tokens from US frontier models since at least late 2024.
Unconfirmed: CISA's claim of Chinese government awareness, and any link to this specific bypass, is not independently confirmed here.
- August mitigations blocked replaying reasoning into a different model.
- The compaction step still ran on the model that produced the reasoning.
- Xint says the attack worked at every reasoning effort it tried.
- CISA named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI.
Impact: Stolen reasoning traces that can aid cheaper model distillation
| Product | Affected |
|---|---|
| OpenAI GPT | gpt-5.6-sol, gpt-4.1 and gpt-4o |
- Vendors
- OpenAI
- Countries
- ChinaUnited States
From the original article
3 min read at securityweek.com
Xint says a flaw in OpenAI’s compaction process let researchers move reasoning traces between models, potentially making it cheaper to distill a rival AI. The post Stolen AI Reasoning Could Help China Build Cheaper Rivals appeared first on SecurityWeek.
By Kevin Townsend · securityweek.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.