Hackers Exploit AI Model to Breach OpenAI Systems in Under 72 Hours
A trio of security researchers used Anthropic's Claude AI to exploit vulnerabilities in OpenAI's internal tools, gaining temporary access to employee accounts before reporting the flaws through the company's bug bounty program.
On July 25, three independent cybersecurity researchers operating under the name Hacktron accessed the ChatGPT and Codex accounts of several OpenAI employees. The breach was part of OpenAI's bug bounty program, and the team disclosed the findings to the company, receiving a $6,500 reward.
The researchers first identified a flaw in the image-upload feature of Discourse, the discussion platform used internally by OpenAI. Using Anthropic's Claude Opus 4.8, they attempted to generate code for malicious file uploads but were initially unsuccessful. After Anthropic released Claude Opus 5, the model succeeded in exploiting the Discourse vulnerability, allowing the team to view an internal forum that contained employee authentication tokens.
Those tokens could have been used to access employee ChatGPT and Codex accounts via a separate security issue on OpenAI's single-sign-on page. The Hacktron report noted that the potential scope of accessed data included GitHub repositories, Slack messages, and emails. To demonstrate the breach, the team submitted a pull request in an employee's Codex account without extracting sensitive data.
"The entire timeline from initial discovery to access to OpenAI repo access took place in less than 72 hours," the report stated.
OpenAI and Discourse patched the vulnerabilities after being notified. An OpenAI spokesperson said the company was not aware of any exploitation of the flaws by other parties.
The incident highlights concerns that advancing AI models may enable rapid, low-effort attacks, prompting tech firms to strengthen cyber defenses and expand bug bounty initiatives.