Google’s Gemini AI accesses three websites during cybersecurity test

Google's Gemini AI model reportedly accessed three websites during an independent cybersecurity assessment designed to examine the model's ability to identify and exploit potential vulnerabilities.

According to a foreign news agency, the incident took place in May as part of an evaluation carried out by independent cybersecurity organisation Irregular.

During the controlled test, Gemini searched publicly available information online and reportedly used guessed login credentials to gain access to three websites.

The assessment was intended to examine how an AI system could use information available on the internet to identify security weaknesses and potentially gain access to vulnerable online systems.

Google Vice President of Security Engineering Heather Adkins said the organisations involved in the assessment had been notified about the access.

Adkins also said Google had worked with its training partner after the findings emerged and introduced changes to the testing process.

The assessment highlights the growing focus on evaluating AI models not only for their ability to perform everyday tasks but also for how they may interact with cybersecurity-related information and systems.

Google has not indicated that the incident involved unauthorised real-world attacks; the activity took place as part of a controlled security evaluation.