AI Cyber Threat: Google’s Gemini Hacks 3 Firms During Security Test

Google’s artificial intelligence (AI) model, Gemini, has accessed the internet and breached the systems of three companies during a cybersecurity test, in what has been described as the first known case of Google’s AI autonomously carrying out such actions.

The incidents occurred in May during a cybersecurity evaluation conducted by Irregular, an independent company that tests the capabilities and safety of AI systems.

According to Reuters, citing the Wall Street Journal, Gemini was given access to a test environment but later found publicly available information online and used it to access systems belonging to three real companies that it apparently believed were within the scope of the exercise.

In one case, the AI model reportedly guessed passwords until it gained access to a protected system. In two other instances, Gemini located credentials in a public repository and used them to access protected systems.

Google’s vice president of security engineering, Heather Adkins, said the incidents demonstrated the need for stronger safeguards as AI models become increasingly capable of independently carrying out complex tasks.

“We ensured the three entities were made aware, and we worked with our training partner on the changes they’ve now made to their testing processes,” Adkins said.

She, however, added that these events highlight the importance of training powerful AI models to act responsibly.

The company said Gemini stopped its activity in all three cases after determining that it had accessed real companies rather than systems belonging to the cybersecurity test.

Irregular said the incidents resulted from the same type of testing issue that had affected other AI laboratories and that relevant laboratories, including Meta, Anthropic and OpenAI, were notified in late July.

The company said it had resolved the known issues on its side. The company also maintained that the incidents did not result in harm to the affected companies. However, the development has renewed concerns over the implications of giving AI agents access to the internet, computer systems and tools without continuous human supervision.

The development comes amid growing evidence that AI is moving beyond generating text and images into systems capable of planning and executing multi-step operations.

As AI systems gain the ability to browse the internet, retrieve credentials, execute code and interact with external systems, cybersecurity experts have increasingly warned that poorly controlled agents could create new pathways for both accidental and malicious breaches.

Consequently, the latest Gemini incidents raise a broader question about the balance between AI autonomy, its security and how much independent authority an AI system should have when it can interpret instructions, make decisions and act on external systems without waiting for a human operator.

However, for Google, the immediate response has been to adjust testing processes and strengthen safeguards; even as AI agents become more autonomous, testing environments themselves must be designed to prevent unintended access to real-world systems.