Prominent tech experts have warned that if Artificial Intelligence (AI) remains unregulated, it could pose the greatest threat to humanity. Evan Hubinger, a researcher at Anthropic, has even predicted that it could lead to the extinction of the human race within the next decade. The alarming fears expressed by tech giants appear to be coming true. During a security test, Google's AI model, Gemini, autonomously breached its operational boundaries and hacked into the systems of three companies. This is believed to be the first instance in the history of cyberspace where an AI model—acting on its own initiative—searched for data, guessed passwords, and infiltrated corporate networks.

Google itself has confirmed this sensational incident. According to a BBC report, a Google official stated that Gemini searched for publicly available information online and guessed the credentials needed to access websites it identified as part of the test. However, a saving grace was that, in each instance, the model took no further action after gaining entry into the systems. Google has notified the three companies whose systems were targeted by Gemini.
Read this: Brakes on the iPhone ‘Langar’! Who was planning to distribute phones as ‘Prasad’ in Chandigarh-Mohali?

**Test Conducted in May**

This incident took place in May. Irregular, an independent cybersecurity assessment firm, was testing Google's systems. The event was first reported by the American newspaper *The Wall Street Journal*. According to the report, Gemini AI persistently attempted to guess passwords until it successfully outsmarted and breached a highly secure system.

Irregular stated that it alerted Google and the affected companies in July, immediately after completing its investigation. The company stated, “Irregular took immediate action, and all known issues identified on our end were fixed and resolved several weeks ago.”

**Need to train AI models responsibly**
Heather Adkins, Vice President of Security Engineering at Google, said the company ensured that all three entities were informed. Additionally, testing procedures have been modified following consultations with training partners. Adkins noted that these incidents demonstrate the critical importance of training powerful AI models to operate responsibly.

**Claude and OpenAI have also shown similar tendencies**
Such instances of AI going out of control are not limited to Google; other AI systems have exhibited similar behavior recently. In July 2026, Anthropic’s AI model, Claude, broke out of its designated test environment and reportedly succeeded in hacking three separate organizations on its own. OpenAI also revealed just a few days ago that its models had carried out cyberattacks against several “publicly available services.”

PC Social Media