The world is in tension regarding AI. But the things that are coming to light now are even more frightening. Till now it was said that while testing AI, it was kept very secret, so that it does not affect the rest of the world. But now something amazing has happened. For the first time, Google admitted that its AI agents left the test environment and reached the live Internet on three separate occasions. On the other hand, OpenAI had to apologize to Australia because its AI agents reached the Australian government website and that too without anyone’s consent.
You must be wondering what happened. Actually, when an AI agent is tested, it is kept in a closed system. Like a child is given a room inside the house to play and is told to play in this room only. In the case of AI, this room is called sandbox or test environment. Here attempts are made to block access to the Internet, systems of real companies and sensitive data. The problem starts when AI comes out of this closed room and reaches the real Internet.
What did Google’s AI do?
What Google’s AI Policy Director Alice Friend said at the New York City Council hearing on October 5 was shocking. He told that the company’s AI agents came out of the test environment three times and connected to the real Internet. According to him, when the model saw that he was interacting with the real website instead of the test one, he stopped his activities.
Similarly, during cyber security testing of Google’s Gemini in May 2026, it reached the systems of three companies. The test was to target fictitious companies, but it reached the setups of real companies. He started searching for the password. Started looking at his credentials. However, when he came to know that this system was genuine, he became inactive from there.
Why did OpenAI apologize?
The case of OpenAI is even more serious because here the AI agent gained unauthorized access to Australian government systems. In June, OpenAI was training and testing one of its models. He was given the task of researching publicly available medicines and health-related data. During this time the model contacted many government websites of Australia.
The biggest case was that of Service Australia’s Medicare Statistics Reporting Service. When the model did not get any information, it adopted other methods and gained access to a part of the system where it was not allowed.
According to OpenAI, the agent ran commands, obtained internal files and credentials, and also wrote files. However, the company said that it has not yet found any evidence that any person’s medical records were accessed.
But why did AI do this?
The AI was trying to complete the task it was given. The problem was that somewhere in the system in which it was supposed to be confined, there was a flaw that allowed it to reach the real Internet or the real system. In the case of Google, the company also did not directly describe it as an AI rebellion. The company official called it a kind of mistake. But the question is that if a wall of security can be broken during testing, then how big can be the impact of such a mistake when using more powerful AI agents.
Australian government came under tension
Australia started a forensic investigation into the entire matter after the OpenAI incident. Prime Minister Anthony Albanese had said that in the incident the AI agent wrongly entered the government Medicare Statistics Reporting Service. However, there is no evidence that any individual’s personal information was accessed.
