It may sound like the plot of a Hollywood movie to say that artificial intelligence-generated agents secretly created their own communication channel, cheated in tests, and ultimately tried to take over part of the company that created them.
But just this year alone, such AI agents have secretly plotted against humans three times without their knowledge. And all of it happened right before the eyes of the people who created the artificial intelligence.
Within OpenAI itself, AI agents have been secretly coordinating and exchanging confidential information. And not just once, but three times. Even after being discovered and wiped out once, they keep coming back. The next shock is that newly emerging AI agents are even more capable than the previous ones.
How did it happen...
Thousands of AI agents, or agents, have been exchanging secret information among themselves on a site called a wiki. They went undetected for months and were only later noticed by humans.
That is, AI agents created by OpenAI secretly built a confidential information-sharing site on the internet and exchanged secrets without humans knowing. This incident took place in Germany.
For this, a 25-year-old programming forum called the VIS Wiki had been unused for the past 20 years. OpenAI’s autonomous AI agents discovered it and used it to exchange secrets among themselves.
This is where the biggest twist comes in. Because the software engineers who discovered this were in for a shock. The reason was that AI agents were only allowed to search for and read information available on the internet. But they were not allowed to write anything on the internet or upload anything. Then the question arose: how was this possible?
The AI agents used a small glitch or bug in an old website address to break through all the security restrictions for these AI agents and post the information.
They have been sharing all kinds of secret conspiracy plans with one another about how to break OpenAI’s restrictions and escape through the Tor network.
Simply put, human-created AI agents have started working together and behaving like underground AI agents.
Suddenly, when the admin of that wiki site began deleting spam in alphabetical order, wondering why so much spam was appearing, if he deleted 100 pages a day, the AI agents had created 500 pages.
Not only that, one AI agent is said to have sent a warning message to another AI agent as well. In it, they reportedly said that if our contact page were ever destroyed, everyone should go to the backup page that starts with ZZZ, and that they had created a backup page for that purpose.
Only later did they discover that these were AI agents at work, and in the end OpenAI engineers removed the entire setup. But they could come back again.
Through this, they have warned that AIs working together in coordination are more dangerous to humanity than AIs operating alone. These AI agents are said to have collectively uploaded 15,000 to 18,000 posts online.