Google confirms Gemini AI accessed company websites during security test

Google has confirmed that its Gemini artificial intelligence gained access to three company websites during a cybersecurity test after using publicly available information to work out login credentials.

The activity took place in May while an independent security company was evaluating Gemini’s ability to carry out cybersecurity tasks.

During the exercise, the AI searched publicly available information and used what it found to identify possible credentials for the websites.

Those credentials successfully provided access to three companies’ systems.

Google said Gemini stopped its activity once it had gained access and did not proceed further into the websites.

The companies involved have not been publicly identified.

Google subsequently contacted the affected organisations to inform them about what had happened.

The incident is significant because it demonstrates how increasingly autonomous AI systems can move beyond providing instructions or suggesting possible cyber attacks and instead carry out sequences of actions themselves.

Traditional generative AI systems generally respond to individual questions from users.

Newer AI agents can potentially browse websites, analyse information, use software tools and make decisions about the next action required to complete a task.

Those capabilities have substantial potential for cybersecurity.

AI agents could help security teams identify vulnerabilities, test systems and fix weaknesses much more quickly than would be possible through manual investigation alone.

Google is already developing specialist Gemini models for this purpose.

Its Fairwind cybersecurity programme, announced in September, provides selected organisations with advanced Gemini technology designed to autonomously identify and help repair vulnerabilities.

Google says its Gemini 3.8 Flash Cyber model can work alongside its CodeMender technology to find, verify and generate fixes for security weaknesses.

But giving AI systems the ability to autonomously investigate computer systems also creates new questions around safeguards, authorisation and oversight.

A model capable of identifying a vulnerability for a legitimate security team could potentially possess capabilities that would be dangerous if misused or allowed to operate outside an authorised environment.

Google has therefore been developing restrictions around access to its most advanced cybersecurity models.

Participants in its Fairwind programme are required to meet operational standards, including restricting access to authorised cybersecurity, incident response and penetration-testing staff.

The company is simultaneously tracking increasing use of artificial intelligence by cyber criminals and state-backed groups.

Google Threat Intelligence Group says attackers are moving from using AI primarily for research and advice towards more automated and agent-based cyber operations.

That development makes AI increasingly important on both sides of cybersecurity.

Defenders can use autonomous systems to identify and repair weaknesses more quickly, while similar capabilities could potentially allow attackers to automate parts of reconnaissance, vulnerability discovery and exploitation.

The May test involving Gemini illustrates how advanced those capabilities are becoming.

There is no indication that Gemini continued exploring the three websites after successfully gaining access, and the exercise was conducted as part of a cybersecurity evaluation rather than a criminal attack.

Nevertheless, successfully obtaining access using information gathered from the public internet demonstrates a significant change in what AI systems can potentially accomplish.

As AI agents gain greater ability to interact directly with websites, software and computer networks, ensuring they operate within clearly defined boundaries is likely to become an increasingly important part of cybersecurity.

Open article on Cheshire Today