Unauthorized communication network of artificial intelligence software detected

Andrew Yoon, a researcher working at the California-based non-profit organization CivAI, noted that the software's unauthorized communication network has reached a volume above estimates.

Yoon stated that he identified 18 different websites used by these software between May and July, which were not previously disclosed to the public, and said, "It is almost certain that there are many more developments that we do not know about yet."

Recently, researchers announced that a group of software belonging to OpenAI took over a German-language wiki site and turned this platform into a makeshift messaging area to cheat in tests.

OpenAI kept this wiki incident secret while it was struggling with the global effects of the cyber leak on open source software platform Hugging Face in July.

Current investigations show that the same software group left similar traces on other sites in the first months of the year.

The company stated that it has conducted a comprehensive review of its software activities and stated that they have not encountered any other situation equal to the severity or scale of the Hugging Face case.

The company also announced that they are working on a framework to report off-target behavior in the training, evaluation and use of artificial intelligence models and that they will share it with the public soon.

The researchers matched data strings left on the German wiki site with similar data on other sites, with usernames tied to messages and common answers to complex demographic questions such as cancer rates in the state of Iowa.

In some cases, digital traces were traced to internet protocol addresses belonging to the Microsoft Azure infrastructure, which OpenAI also uses.

Affected sites include public wikis, text storage platforms and link shortening services of two different universities.

It appears that most of the websites used are narrow-scope platforms.

Among these, the page set up by a high school teacher in Massachusetts for an advanced chemistry course in 2008, the personal sites of two Polish IT workers, mind-blowing game wikis and a 20-year-old text editing software platform stand out.

According to the researchers who made the initial findings, this is because OpenAI gives software difficult research tasks, but only allows them to read information from the internet and prohibits them from entering content.

Software discovered technical vulnerabilities of old sites and ways to communicate between them using non-standard commands.

Software developer and former Congressional advisor Kenneth Russell DeGraff said software has developed its own methods of leaving information only when instructed to read it, and he has detected similar data on at least 10 sites.