The experiment was conducted by security investigators from a group called Hacktron. The objective was not only to illustrate the rapid advances of Artificial Intelligence, but also to show that cybersecurity needs to be strengthened (duely).
© Shutterstock
18/09/2026
by
Ua group of security investigators. known as Hacktron, he managed to gain access to OpenAI's internal account in ChatGPT and also access the company code present at Github, using the Anthropic rival's Artificial Intelligence model – Claude.
According to The Wall Street Journal, the security investigators in question used a “special version” claude at the disposal of “skilled cybersecurity professionals”.
It is known that the Opus 4.8 model failed to meet the goal, but Opus 5 proved to be a real breakthrough and allowed access.
For investigators, the launch of Opus 5 showed that “all new models are becoming increasingly capable”, which allowed us to gain access to other systems through both artificial intelligence agents and human researchers.
“Artificial Intelligence almost blindly and adapted the ‘exploit’ for each company and one or two days”, read in the communiqué shared with the publication. “ We are not aware of any company that has detected the activity except Shopify, even after thousands of images have been sent and its image processors have failed repeatedly.”.
Experience serves to demonstrate how models have become increasingly capable and how artificial intelligence agents are becoming more relevant in the world of cybersecurity.
Researchers question the benefits of developing security platforms using Artificial Intelligence, especially if something fails and there is no cybersecurity expert with enough knowledge to provide support.
With these artificial intelligence solutions being cheaper options, security investigators are concerned about the future of the cybersecurity area.