Three Indian-origin cybersecurity researchers from Hacktron AI—Harsh Jaiswal, Mohan Pedhapati and Rahul Maini—used Anthropic’s Claude to exploit vulnerabilities in OpenAI’s systems under OpenAI’s bug-bounty programme, gaining access to several OpenAI employees’ ChatGPT and Codex accounts and reaching the company’s private GitHub environment. The investigation began with a vulnerability in OpenAI’s public community forum, hosted on the Discourse platform, where a flaw involving the processing of certain image files could potentially allow code execution on the forum’s server. Claude helped investigate the vulnerability and develop an exploit, assisting with writing, debugging and adapting code; after a newer Claude model became available, it reportedly produced a successful exploit much faster. The researchers then found another weakness involving authentication tokens, gaining access to multiple employee accounts, one connected to OpenAI’s private GitHub environment. They created a harmless pull request modifying a documentation file, which was not accepted, and stopped testing after establishing that sensitive internal material could potentially be accessed. They said they did not read or download OpenAI’s private source code. OpenAI fixed the vulnerabilities and revoked affected authentication tokens and sessions.