Researchers Used Claude to Hack OpenAI and Gain Access to Secret Repository Researchers at Hacktron AI used Claude to exploit a bug in the Discourse forum software OpenAI relies on for its community platform. Early attempts failed — but a day after Anthropic shipped Claude Opus 5, the model found a working exploit. The chain led to stolen auth tokens, then GitHub access, then read access to OpenAI's internal Monorepo (no model weights exposed). The researchers stopped once they realized the access was real, disclosed responsibly, and got paid $6,500 under OpenAI's bug bounty. The bigger story: this wasn't a nation-state team. As Hacktron's CTO put it, they're "three guys with Claude and Codex subscriptions." OpenAI took the incident seriously enough to pull 25% of its production engineers off other work to hunt for more holes. AI isn't just writing exploits faster — it's collapsing the skill floor needed to find them at all. #AI #Cybersecurity #Claude #OpenAI #InfoSec #Anthropic #Web3