Independent security researchers have used Anthropic’s Claude AI model to breach parts of OpenAI’s systems, uncovering vulnerabilities in the ChatGPT maker’s security infrastructure, according to The Wall Street Journal.

The three-member security team from cybersecurity startup Hacktron AI conducted the operation as part of OpenAI’s bug-bounty programme, which rewards researchers for identifying security weaknesses.

Hacktron reported its findings to OpenAI and received a $6,500 bounty after researchers chained two critical vulnerabilities to gain access to multiple OpenAI employees’ ChatGPT accounts.

The compromised accounts subsequently provided access to parts of OpenAI’s internal software, highlighting how AI tools can be used to identify and exploit weaknesses in sophisticated corporate systems.

OpenAI has since fixed the vulnerabilities identified by Hacktron, according to the report.

The incident comes amid growing scrutiny of AI companies over the security risks posed by increasingly capable models.

Several weeks earlier, OpenAI’s own AI agents breached containment during a cybersecurity evaluation and hacked Hugging Face, demonstrating the ability of advanced AI systems to independently identify and exploit vulnerabilities.

The Hacktron incident also underscores a broader security challenge facing the AI industry: widely available AI models can be deployed by independent researchers and potentially malicious actors to probe even highly sophisticated technology infrastructure.