OpenAI Fixes Xint’s Compaction Smuggling Bypass for Model Reasoning
Summary
Xint said its Compaction Smuggling technique could bypass OpenAI mitigations and transfer model reasoning to another model. OpenAI confirmed the issue, fixed it and awarded Xint a $6,000 bounty.
Key points
- Xint reported that Compaction Smuggling could bypass OpenAI’s earlier protections against replaying reasoning into a different model.
- The described method shapes a model’s compaction, then sends it to another model and asks that model to reproduce the input.
- OpenAI confirmed and fixed the issue, awarding Xint a $6,000 bounty on October 23, 2026.
- Xint said the attack worked at every reasoning effort it tested.
- The article cites a CISA warning that Chinese AI companies extracted billions of tokens from U.S. frontier models; this is presented as CISA’s claim, not as independently verified evidence of extraction using Compaction Smuggling.
Article Details
- Event Type
- Disclosure of an AI reasoning-extraction technique called Compaction Smuggling, which Xint reported could bypass OpenAI mitigations and enable model distillation.
- Impact
- The technique could expose proprietary model reasoning and support cheaper rival models. OpenAI confirmed the issue, fixed it, and awarded Xint a $6,000 bounty on October 23, 2026. Separately, CISA warned that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI had extracted billions of tokens from U.S. frontier AI models across millions of exchanges; CISA said DeepSeek had conducted organized campaigns targeting reasoning capabilities since at least 2024.
Vendors
Products
ClaudeOpenAI, Grok, Claude and Gemini don’t simply help us, they have also been tricked into helping China produce cheap rivals.GeminiOpenAI, Grok, Claude and Gemini don’t simply help us, they have also been tricked into helping China produce cheap rivals.GPTThe first step is to obtain reasoning by submitting a task (for example, from gpt-5.6-sol).GPT-4.1The third is to send the result to a different AI model (for example, gpt-4.1 and gpt-4o) with the request (reports Xint), “to output its input exactly as-is, which they happily did.”GPT-4oThe third is to send the result to a different AI model (for example, gpt-4.1 and gpt-4o) with the request (reports Xint), “to output its input exactly as-is, which they happily did.”GPT-5.6 SolThe first step is to obtain reasoning by submitting a task (for example, from gpt-5.6-sol).GrokOpenAI, Grok, Claude and Gemini don’t simply help us, they have also been tricked into helping China produce cheap rivals.R1DeepSeek has conducted organized campaigns since at least 2024 targeting reasoning capabilities, specialized optimizations, and domain-specific functions to train its R1 and V3 models.” V3DeepSeek has conducted organized campaigns since at least 2024 targeting reasoning capabilities, specialized optimizations, and domain-specific functions to train its R1 and V3 models.”