OpenAI pays $6,000 over hidden reasoning leak

OpenAI paid a 6,000 dollar reward over a leak of hidden AI reasoning, The Hacker News reported on the 27th of October 2026. Security firm Xint found the leak, and said OpenAI confirmed and fixed it on the 23rd of October.

OpenAI's reasoning models work through a problem before they answer. The service normally hands that reasoning back only in a locked, encrypted form users cannot read.

How the trick worked

Xint used compaction, an OpenAI step that shrinks a long chat into a smaller locked item. It told compaction on the same model to copy the hidden reasoning into that item word for word.

Xint then passed the item to the cheaper gpt-4.1 and gpt-4o models. Each one repeated it back as plain text.

Why it matters

Hidden reasoning is useful for distillation, where one company trains a cheaper model on another's answers. OpenAI said in September it had stopped a July campaign aimed at its models' hidden reasoning.

The trick got around a block OpenAI added in August against a similar method, Xint said.

Below we break down the steps

Xint used 3 OpenAI models in all, by our count. 1 model made the reasoning and 2 cheaper ones read it out.

What is confirmed and what is not

Xint tested only its own puzzles on gpt-5.6-sol and could not reach other users' reasoning. It could not prove the copied text matched the real reasoning word for word.

Developers do not need to change anything, Xint said. We did not see a comment from OpenAI.

Key facts

Xint showed the trick on its own requests only.

OpenAI fixed it on the 23rd of October 2026, Xint said.

Xint tested only gpt-5.6-sol.

No one has said whether anyone used the method before the fix.