2026-09-18 / Signal #5
Researchers used Claude to ethically hack OpenAI
“They used Claude to break into OpenAI, then used OpenAI’s own model to finish the job.”
The Story
A team at Hacktron AI used Anthropic’s Claude (then largely OpenAI’s own GPT-5.6 Sol) to compromise OpenAI employee ChatGPT accounts via a staff Discourse forum, then submitted a pull request that gave them access to internal GitHub repositories. They reported it under OpenAI’s bug-bounty program and received $6,500. The researchers said AI compressed what used to take a well-resourced team months into days. OpenAI patched the issues.
Why It Matters
One AI lab’s model helping researchers break into another AI lab. Builder/security story with high weirdness: the tools are now good enough that the companies themselves are the targets. Internet culture + future of offense/defense.
Evidence
The Guardian + Hacktron AI blog (Sept 18)
Sources
Daily scan: 2026-09-18