OpenAI Finds a Reasoning Thieves’ Ring — But Leaves Azure Vulnerable
An army of 15,000 accounts was taken down for scraping OpenAI’s hidden reasoning. The attack was repeated on Microsoft Azure to scrape hidden reasoning from GPT-6 Astra and Anthropic’s models.
OpenAI discovered a network of over 15,000 accounts on their platform. These folks were extracting the encrypted reasoning that the company hides from users from their models during the summer. At their peak, on July 24–25, they made 16,000 requests from 4,000 accounts, and OpenAI believes some members of this network are connected to those involved with Moonshot AI who created the Kimi model. OpenAI patched this issue on their end by July 28. However, on September 13 researchers discovered that they could continue using this method on Microsoft Azure to access GPT-6 Astra and Anthropic’s models. They only fixed this issue on September 27 and 28.
I think what’s interesting about this story is how vulnerable a model can be when it’s hosted on a third party platform. For example, OpenAI patched this issue on their end by July 28 but the flaw was still working on Azure on September 13, and it wasn’t patched there until September 27. In my own agent setups I always check exactly which API and infrastructure a model is actually running through, not just the model version — this case is one more reason why. You should never assume that if you see “the same model” on different platforms that it will be equally secure.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles