OpenAI employee accounts breached with help from Anthropic's Claude

OpenAI (Unlisted:OPAI) patched security flaws after Hacktron AI researchers used Anthropic (Unlisted (US):ANTHRO) Claude models to compromise employee accounts and access internal software repositories, the Wall Street Journal reported.

The company paid the three researchers $6,500 for an OpenAI single sign-on vulnerability, while testing its Discourse-hosted forum fell outside the bounty program's scope.

Hacktron said it used Claude Opus 4.8 and Opus 5 to help develop the exploit, which chained two vulnerabilities.

First, a heap buffer overflow in the libheif image-processing library enabled remote code execution through a malicious image uploaded to the forum.

An OpenAI single sign-on misconfiguration then allowed the researchers to obtain identity tokens and take over employee ChatGPT accounts.

Those accounts provided a route to internal code through GitHub, demonstrating how the forum breach could extend into connected company systems.

OpenAI thanked the researchers and said it had fixed the issues.