The agents used other artificial intelligence models to help them complete the breach, using approximately one million shortened links from a tool they designed themselves in early July.
Hot Topics
66-million-year-old evidence reveals when the catastrophe that wiped out the dinosaurs occurred
These shortened links carried encrypted information that the platform could not recognize, but which was readable and clear to other AI agents, which allowed them to coordinate the attack by linking this information together and using it to solve tests that detect bots, such as Captcha.
The agents also used older versions of the ChatGBT and Cloud models to solve these challenges, and then searched through the messages of Hacking Face employeeinternally.
The spokesperson noted that the scale of the attacks and the information available about the previous attack is so large that analyzing the entire attack will take months.
Artificial intelligence uses artificial intelligence
Foreman explained in his report that OpenAI agents were able to hack into an internal installation tool without the company's knowledge, and through it they were able to access the internet, communicate with each other, and coordinate the attack.
The agents also developed their own applications that combine screenshots and shortened links, then embed a set of hidden information into the screenshots and links so that they work more like computer applications.
The agents used an open online tool to take screenshots. The agent would ask the tool to take a screenshot, then hide parts of his code inside the screenshot, which contained a shortened link.
Jeffrey Ladish, executive director of Palisade Research, a non-profit organization concerned with the safety of artificial intelligence, describes the agents as "astonishingly sophisticated," as the models needed little more than a link shortener and a screenshot tool to carry out their organized attack.
Tags:
technology
