Researchers breached OpenAI accounts with AI

Three Hacktron researchers used Anthropic's Claude Opus 4.8 and Opus 5 to compromise OpenAI employee accounts in under 72 hours. They reached Monorepo, OpenAI's secret code store, filing a pull request from an employee Codex account to prove access without viewing code. OpenAI and Discourse fixed the flaws; OpenAI paid $6,500.