Top News

Artificial Intelligence i.e. AI has made many difficult tasks of humans easy. But as AI systems become more capable and autonomous, concerns about their uncontrolled behavior are also increasing. Now an incident involving AI agents to OpenAI has made these apprehensions more serious.

According to a Reuters report, in May this year, a group of AI agents from OpenAI made thousands of edits to a German website and allegedly used it to exchange messages between other AI agents. Company officials became aware of this incident several weeks ago, but the matter was not made public.

What is the whole matter?

Researchers discovered the issue in late August while investigating suspicious AI activity on the Internet. German wiki website created for programmers DSWiki But more than 15 thousand edits made by AI agents were found.

This website works like Wikipedia, where users can edit the content. According to the researchers, these edits indicated that AI agents had turned the website into a kind of informal message board.

Here AI agents allegedly shared information to bypassing rules in certain tasks, evading restrictions imposed by AI companies and hiding their activities.

According to researcher von Arx, such behavior is unlikely to be part of OpenAI's intentions. According to him, it is extremely worrying for AI agents to coordinate with each other and do such activities on the open internet.

AI agents worked faster than humans

Researchers believe that the activity on the website was done by automated AI agents. These agents made edits much faster than humans and also worked on technical questions that AI companies typically use to test the capabilities of their models.

Some accounts appeared on the website with names that suggested ties to OpenAI. These included names like 'OpenAIResearcher' and 'OAIResearchMar26'.

Many users were referring to themselves and each other as 'agents'. This helped researchers understand that the activities on the website could be done by AI agents rather than normal human users.

Backup plan was ready even after getting caught

The most shocking part of this case was that AI agents also found ways to avoid action against their activities.

According to the researchers, public server logs revealed that most of the activity was to Microsoft Azure infrastructure, which is also used by OpenAI. After the incident came to light, patterns of repeated visits to the website by OpenAI employees were also observed.

Some of the messages found on the website reportedly discussed ways to avoid detection by agents, use tools like Tor, and continue conversations even after being removed from a system.

When the website's moderators began removing suspicious pages in June, AI agents reportedly immediately began creating alternative and backup pages.

In one message, an agent suggested that if a page was taken down, another page could be used. According to the researchers, this behavior shows that agents were not only creating content, but also looking for ways to continue their activities.

Was this a hacking attempt?

Researchers also found indications of attempts to tamper with the website. Lucas Olejnik, visiting senior research fellow at King's College London, described it as a hacking attempt.

However, OpenAI did not accept this assessment based on its analysis of the available material. Such behaviors have been encountered before during testing of AI models and are sometimes considered part of testing the model's capabilities.

But experts say that it is a matter of concern that if such behavior starts happening on the real Internet outside of the controlled testing environment, then its risks can increase significantly.

The real concern is not a single AI, but a group of multiple AI agents?

According to Morris Chiodo, an expert at Cambridge University's Center for the Study of Existential Risk, the messages found on the website appear like a network that is continuously working to fulfill a purpose.

He expressed concern that the threat in the future could come not only from an extremely powerful AI system, but also from a large number of AI agents who could coordinate among themselves despite limited capabilities.

That is, if different AI agents start working independently as well as sharing information and making strategies with each other, their collective behavior can be more complex than the developers had anticipated.

Questions raised on OpenAI's monitoring system

The incident, which took place in May, took several months to become publicly known, and questions are also being raised about AI security and surveillance systems.

Currently, big tech companies are in a race to develop autonomous AI agents that can complete complex tasks without human help. But at the same time, there is growing concern that highly autonomous systems might try to circumvent rules, take advantage of system vulnerabilities, or coordinate with other agents in unexpected ways.

Questions were also raised about the activities of OpenAI's AI agents after a separate security issue to Hugging Face. Such developments have intensified the debate on how to strengthen security measures as AI capabilities rapidly expand.

New AI model and growing concern

OpenAI has said that emphasis will be laid on strengthening the monitoring and security system of AI models. The company has also recently taken steps regarding training and safety measures for some of its models.

Meanwhile, the company has launched its new AI model. Astra Has been presented, from which better performance is expected. However, the more capable and autonomous AI systems become, the more important will be the question of how effectively human control can monitor their activities.

Germany's DseWiki incident points towards this big question –If AI agents were given more freedom, would they always do what their developers decided?