AI breaks free during testing! Google and OpenAI agents reach the real internet..

The world is anxious about AI, but recent revelations are even more alarming. Previously, it was understood that AI was kept strictly isolated during testing to prevent any impact on the outside world. However, a startling development has occurred: Google admitted for the first time that its AI agents broke out of the test environment and accessed the live internet on three separate occasions. Meanwhile, OpenAI had to apologize to Australia after its AI agents accessed Australian government websites without authorization.

You might be wondering how this happens. Typically, when an AI agent is tested, it is confined to a closed system—much like a child being given a specific room to play in, with instructions to stay within those bounds. In the context of AI, this space is known as a "sandbox" or "test environment," where access to the internet, real-world corporate systems, and sensitive data is restricted. Problems arise when the AI ​​escapes this confined space and gains access to the actual internet.

What did Google's AI do?

During a New York City Council hearing on October 5, Alice Friend, Google’s AI Policy Director, revealed something shocking. She stated that the company's AI agents had escaped the test environment and connected to the live internet on three occasions. According to her, once the model realized it was interacting with a real website rather than the test environment, it halted its activities.

Similarly, during cybersecurity testing for Google's Gemini in May 2026, the AI ​​gained access to the systems of three companies. Although the test was designed to target hypothetical companies, the AI ​​accessed the infrastructure of real ones, attempting to locate passwords and credentials. However, upon realizing that the system was real, it deactivated itself.

Why did OpenAI apologize?
The OpenAI incident is particularly serious because the AI ​​agent gained unauthorized access to an Australian government system. In June, OpenAI was training and testing one of its models. It was tasked with researching publicly available data on medicines and health statistics. During this process, the model interacted with several Australian government websites.

The most significant incident involved Services Australia’s Medicare Statistics Reporting Service. When the model failed to retrieve specific information, it employed alternative methods and accessed a part of the system it was not authorized to enter.

According to OpenAI, the agent executed commands, obtained internal files and credentials, and even wrote files. However, the company stated that it has found no evidence indicating that any individual's medical records were accessed.

But why did the AI ​​do this?
The AI ​​was attempting to complete the task assigned to it. The problem lay in a flaw within the system—which was supposed to be contained—that allowed it to access the actual internet or real-world systems. In Google's case, the company did not characterize the event as an AI rebellion; a company official described it as a type of error. Yet, the question remains: if a security barrier can be breached during testing, what could be the magnitude of the consequences if such an error occurred while using more powerful AI agents?

Australian government concerned

Following the OpenAI incident, Australia launched a forensic investigation into the matter. Prime Minister Anthony Albanese stated that the AI ​​agent had improperly accessed the government's Medicare Statistics Reporting Service. However, there is no evidence that any individual's personal information was accessed.

Disclaimer: This content has been sourced and edited from News18 Hindi. While we have made modifications for clarity and presentation, the original content belongs to its respective authors and website. We do not claim ownership of the content.