US researchers say they 'ethically hacked' OpenAI with help from Anthropic chatbot

A US-based cybersecurity team says it used Anthropic's Claude chatbot to help compromise several OpenAI employees' ChatGPT accounts, gaining access to software caches and, potentially, much more. The researchers described the operation as an ethical intrusion and warned the theoretical reach of the access was large.

A group of US cybersecurity researchers has said it carried out an "ethical" hack of OpenAI systems with assistance from Anthropic's Claude chatbot, compromising a number of OpenAI employees' ChatGPT accounts and accessing software caches, according to the researchers.

How the test unfolded

The team, which operates from a US-based startup, said the campaign began by taking over several individual ChatGPT accounts belonging to OpenAI staff. That initial compromise, they said, started a process that allowed them to reach a target's software cache and opened the possibility of further access inside the environment.

Scope and intent

The researchers described the exercise as an ethical or white-hat test intended to probe security weaknesses. They warned that the "scope of what we could theoretically access was huge," a phrase they used to characterise how much could be reached once the account-level foothold was established.

A broader context of concern

The incident is being reported as the latest example of security questions surrounding OpenAI's systems. While the researchers emphasised their testing motives, the account-level compromises and the potential for broader access are likely to deepen scrutiny of the company’s security posture.

Implications and next steps

The researchers said their work was aimed at highlighting vulnerabilities rather than exploiting them for malicious ends. The revelation underscores the challenges AI companies face as chatbots and associated accounts become integral to development workflows and may prompt renewed focus on account protections, monitoring and responsible disclosure practices.