[Digital Today reporter Chi-gyu Hwang] Hacktron AI's security research team showed it could access an OpenAI employee's ChatGPT account using Anthropic's AI Claude, the Wall Street Journal reported on Sept. 17 local time.

The team found the issue while participating in OpenAI's bug bounty program and reported it immediately, the report said. OpenAI paid a $6,500 reward, and the team disclosed the results for the first time through the WSJ.

Mohann Peddhapati (모한 페드하파티), Hacktron AI's chief technology officer, said, "We do not have as many resources as hacking groups backed by the Chinese government." He said, "We are just 3 people who only pay for Claude and Codex subscriptions."

OpenAI said it found two issues, a vulnerability in the external service Discourse that runs its community forum and a problem within OpenAI itself. It said it narrowed the permissions of community login tokens and revoked affected tokens and sessions.

The attack began on July 23 when Hacktron AI's research team found a bug in Discourse's image file processing. The team requested attack code from a special version of Claude Opus 4.8 but initially failed. Anthropic released Opus 5 that evening, and the next day Claude found a way to exploit the bug. Using that code, the team accessed Discourse servers and obtained authentication tokens, it said.

Some tokens belonged to OpenAI employees and were also valid for ChatGPT and GitHub. The team said the repository it accessed was named "monorepo". Monorepo was described as a large repository containing OpenAI algorithm secrets, but it was said not to include model weights.