WSJ: OpenAI systems hacked using the AI model Claude from Anthropic

Independent cybersecurity researchers used Claude software from Anthropic to access an OpenAI employee account on the ChatGPT service. This allowed them to view the company’s closed code repository and suggest changes to it, according to The Wall Street Journal.

The procedure took place as part of OpenAI’s special program to find vulnerabilities in its systems. A group of researchers promptly informed the company of the results of their work. OpenAI paid the team $6,500, and the researchers told the WSJ about their work.

OpenAI reported that the researchers found two vulnerabilities: one in the third-party service Discourse, which runs the OpenAI community forum, and the second in the systems of the company itself. According to OpenAI, both problems have been resolved. “We thank the researchers for contacting us and sharing their findings. We limited access rights, as well as canceled the compromised tokens and completed the relevant sessions, the company said

Anthropic does not comment on this information.

Eugene Hvostik