AI came out of human hands! Google Gemini discovered passwords on its own and hacked 3 companies

New Delhi. Veteran tech experts have warned that if Artificial Intelligence (AI) is not controlled, it can prove to be the biggest threat to humanity. Anthropic researcher Evan Hubinger has predicted the extinction of the human race in the next decade. The worst fears of the tech giants seem to be coming true. Google’s AI model Gemini (Gemini AI) crossed its limits and autonomously hacked the systems of three companies during a security test. This is considered to be perhaps the first such case in the history of cyber world, where an AI model at its own discretion searched the data, guessed the password and broke into the companies’ networks.

This sensational case has been confirmed by Google itself. According to a BBC report, a Google official said that Gemini searched public information available online and guessed the credentials to access websites, which it thought was part of the test. However, it is a matter of relief that in every case the model did not take any further action after entering the system. Google gave information about the three companies whose systems were targeted by Gemini.

The test was done in May

This matter is of the month of May. Irregular, an independent company that evaluates cyber security, was testing Google’s system. This incident was first reported by the American newspaper Wall Street Journal. According to the newspaper report, Gemini AI continued to guess the password and eventually it even cracked a highly secure system and broke into it.

On this matter, Irregular has said that after completing its investigation, it had alerted Google and the affected companies in July itself. “Irregular took immediate action and all known issues we encountered were fixed and resolved several weeks ago,” the company said.

Need to train AI models responsibly

Heather Adkins, Vice President of Security Engineering at Google, said that the company ensured that all three organizations were informed about this. Also, changes have been made in the testing procedures after discussing with the training partner. Adkins said these incidents are proof of how important it is for powerful AI models to be trained to act responsibly.

Claude and OpenAI have also shown their attitude

This uncontrolled AI is not limited to Google only. In recent times, other AI systems have committed similar misdeeds. In July 2026, Anthropic’s AI model Claude broke out of its default test environment and reportedly managed to hack three different organizations on its own. OpenAI had also revealed a few days ago that its models had carried out cyber attacks against several “publicly available services”.