Researchers breach OpenAI's systems using Anthropic's Claude chatbot

Cybersecurity researchers have hacked into OpenAI's internal systems with the help of Anthropic's Claude chatbot, compromising employee accounts and source code.

• Hacktron AI researchers compromised multiple OpenAI employees' ChatGPT accounts through the staff's Discourse discussion forum, gaining access to internal software cache and GitHub code.
• Researchers said the hack used Claude to generate exploit code and largely ran on OpenAI's own GPT-5.6 Sol model, taking a few days at a cost under $3,000.
• OpenAI paid the team $6,500 under its bug bounty program; Hacktron said it accessed but did not download OpenAI's repository code.
• An OpenAI spokesperson thanked the researchers and said the company had addressed the exploited vulnerabilities.
• The researchers exploited two linked vulnerabilities in OpenAI's Discourse forum, gaining access to ChatGPT and Codex accounts.
• Hacktron completed the attack chain in under 72 hours in July, crafting the exploit image using Anthropic's Claude models.