So the mechanics: Hacktron AI used Claude to write the attack code, entered through OpenAI's staff discussion forum on Discourse, compromised several employees' ChatGPT accounts, and then filed a pull request against OpenAI's GitHub service. That pull request was the pivot — it exposed a software cache, and the researchers say the theoretical blast radius beyond it was "huge." They had access to repository code but did not download it, reported under OpenAI's bug bounty program, and collected $6,500. Two details worth noting for practitioners. First, the attackers say most of the heavy lifting was done not by Claude but by OpenAI's own GPT-5.6 Sol — the target supplied the crowbar. Second, the time compression claim: work that took a well-resourced team months now compresses into days. That's the operational takeaway, and it applies to your systems, not just OpenAI's. Dario presumably has a 15,000-word essay about this, and a wet lab. The lab is for the wet labs. OpenAI thanked the researchers, patched the holes, and warned the pace can't continue at maximum speed — while presumably continuing at maximum speed.