OpenAI Says It Caught a Massive AI Heist: 15,000 Accounts Were Trying to Steal Its Models' Hidden Reasoning

OpenAI says it just broke up one of the most brazen attempts to steal AI capabilities ever seen — a coordinated operation of more than 15,000 accounts systematically extracting the hidden reasoning its models use to think.

The disclosure came on October 1, 2026, when the company said it had "identified and disrupted a coordinated campaign designed to extract protected reasoning from our models." The activity, which OpenAI calls adversarial distillation, is the systematic and unauthorized use of one model's outputs or reasoning to help develop another model.

What the operators did

The campaign started quietly on July 1 at low volume, then ramped sharply. On July 24 and 25 alone, OpenAI observed 16,000 requests using an extraction pattern from more than 4,000 users. Follow-up investigation found related prompt-pattern activity across a cluster of over 15,000 users — all of it, OpenAI says, fully disrupted by July 28.

Here's the thing: nobody broke OpenAI's encryption or compromised a database. The operators manipulated their interactions with the models so that protected reasoning — the model's internal record for working through a task, the inner monologue you never see — was reproduced in a form visible to the requester.

One method was especially sneaky: operators copied encrypted reasoning from one conversation and asked the model in a different conversation to decrypt and transcribe it. OpenAI also confirmed attack paths reported by independent security researchers through responsible disclosure were real, and closed a pathway that could let someone possessing another user's encrypted reasoning replay it to recover its contents.

Who was behind it

OpenAI said it was unclear whether all the operators originated from a single actor. But it attributed a core cluster of the activity to individuals associated with Moonshot AI, the company behind the Kimi assistant.

Why it matters

Extracted reasoning reveals material deliberately withheld from the final answer, and — this is the critical part — it can be used to train another model without preserving the safeguards applied to the original. OpenAI framed this as both a safety and a national security risk: as models gain capabilities in dual-use areas, the risks of this kind of theft grow. The company expects adversarial distillation attempts to become more sophisticated as frontier models improve, because stealing intelligence is far cheaper than training it.

AiPost's take: the AI arms race just opened an espionage front. Instead of spending billions to train frontier models, some actors are trying to siphon the thinking out of others' models. Protecting hidden reasoning is about to become as important as prompt safety — and this probably isn't the last heist we hear about.