Xint says an OpenAI flaw let researchers move reasoning traces

Xint says a flaw in the compaction process of OpenAI let researchers move reasoning traces between models. Reasoning traces are the step by step notes that a model writes while it works on a problem. Xint says this could cut the cost to distill a rival AI. To distill is to train a cheaper model from the output of a stronger one. The report gives no number and no test result.

What Xint says

Xint says the flaw is in a compaction process at OpenAI. With the flaw, researchers could move reasoning traces from one model to a different model. The report does not tell how compaction works.

The proof

The report gives no test results, no cost figure and no data on the method. It does not show that a person used the flaw to make a rival model.

What is not known

The report does not say if OpenAI agrees with Xint or has repaired the flaw. It does not say who could use it. The lower cost is a possible result that Xint gives, not a measured result.

Insight: Rules guard model weights, but traces may also teach a rival

On 16 October 2026 China published a draft of export controls. It names model weights and AI agents, but API access stays free. Xint points at one more asset: reasoning traces. This looks like a different risk: rules name the weights, while the traces can be what teaches a copy. Leaders must tell vendors to show what their models give out, and not only what they keep.