Hacktron AI Researchers Used Anthropic Claude to Reach OpenAI Internal Repositories

Security researchers at Hacktron AI say they used Anthropic’s Claude models to turn a Discourse image-processing flaw and an OpenAI single sign-on misconfiguration into access to OpenAI employees’ ChatGPT and Codex accounts, including a path to an internal repository. OpenAI fixed the OpenAI-side issue within roughly 14 hours of submission and later paid a $6,500 bounty.