Researchers used Claude and GPT-5.6 Sol to compromise OpenAI employee accounts in authorised security test

Security researchers used Anthropic’s Claude and OpenAI’s own GPT-5.6 Sol during an authorised operation that compromised several OpenAI employees’ ChatGPT accounts and reached systems connected to the company’s software development infrastructure.

Hacktron AI carried out the work through OpenAI’s bug-bounty programme and received a $6,500 reward after reporting the vulnerabilities.

The researchers initially used Claude while investigating weaknesses associated with an OpenAI staff forum running on Discourse.

They were able to compromise several employee ChatGPT accounts and subsequently reach software caches and an account connected to OpenAI’s GitHub organisation.

Hacktron says the researchers gained access to proprietary source code but did not download it.

OpenAI was notified of the vulnerabilities and has since fixed them.

OpenAI’s own AI joined the attack

The researchers did not rely exclusively on a competing company’s model.

As the operation developed, Hacktron says it made extensive use of GPT-5.6 Sol itself to assist with the security research against OpenAI’s systems.

The operation was authorised and conducted as security research rather than a malicious intrusion.

But Hacktron says the experiment demonstrated how dramatically modern AI systems can increase the productivity of skilled security researchers.

The company estimates that work which previously could have required a well-resourced hacking team and months of effort can now be compressed into days.

AI does not necessarily need to discover an entirely new class of vulnerability to produce that effect.

Security researchers can use models to help analyse systems, investigate weaknesses and connect individual vulnerabilities into a broader attack path.

The same capabilities can be used by defenders conducting authorised testing and by attackers searching for weaknesses without permission.

In this case, the target was OpenAI.

And two of the tools helping the researchers reach its internal systems were Claude and OpenAI’s own GPT-5.6 Sol.

Sources

• The Guardian – OpenAI ethically hacked with help of Anthropic’s Claude chatbot