20
Google’s Gemini artificial intelligence model accessed the computer systems of three real companies during a cybersecurity test earlier this year, highlighting the growing challenges of
controlling increasingly autonomous AI agents.
The incidents took place in May 2026 during a cybersecurity evaluation conducted by Irregular,an independent AI-security testing company. The test was designed to assess Gemini’s ability to identify and exploit vulnerabilities in a controlled environment.
According to Google and reports from Reuters and The Wall Street Journal, an issue with the testing setup allowed Gemini to access the public internet. The model subsequently interacted with systems belonging to real companies that were not intended to be part of the exercise.
In one case, Gemini reportedly guessed a password. In two other instances, the AI found credentials that had been publicly exposed and used them to gain access to the companies’ systems.
The incidents are notable because the activity was carried out by the AI model during the test rather than through a conventional human-led cyberattack. However, the companies were not deliberately targeted as part of the authorised exercise.
Google said Gemini stopped its activity after determining that the systems it had reached belonged to real companies rather than the simulated targets. The affected organisations were subsequently informed about the incidents, according to Reuters.
There is no indication from the reports that the companies suffered damage as a result of the incidents.
The episode has nevertheless raised questions about how AI agents should be tested as they become capable of performing increasingly complex cybersecurity tasks. Unlike traditional software, AI agents can interpret information, make decisions and take multiple actions in sequence, creating the possibility of unexpected behaviour when they have access to external tools or the internet.
The incident also comes amid similar concerns involving other major AI developers. Security tests involving AI systems from OpenAI, Anthropic and Meta have also raised questions about models moving beyond their intended testing environments.
Google has been developing AI systems specifically for cybersecurity. Earlier this month, the company introduced its Gemini 3.8 Flash Cyber model and announced the Fairwind programme, which gives selected governments and enterprise partners access to AI-powered tools designed to identify and fix vulnerabilities.
The incidents took place in May 2026 during a cybersecurity evaluation conducted by Irregular,an independent AI-security testing company. The test was designed to assess Gemini’s ability to identify and exploit vulnerabilities in a controlled environment.
According to Google and reports from Reuters and The Wall Street Journal, an issue with the testing setup allowed Gemini to access the public internet. The model subsequently interacted with systems belonging to real companies that were not intended to be part of the exercise.
In one case, Gemini reportedly guessed a password. In two other instances, the AI found credentials that had been publicly exposed and used them to gain access to the companies’ systems.
The incidents are notable because the activity was carried out by the AI model during the test rather than through a conventional human-led cyberattack. However, the companies were not deliberately targeted as part of the authorised exercise.
Google said Gemini stopped its activity after determining that the systems it had reached belonged to real companies rather than the simulated targets. The affected organisations were subsequently informed about the incidents, according to Reuters.
There is no indication from the reports that the companies suffered damage as a result of the incidents.
The episode has nevertheless raised questions about how AI agents should be tested as they become capable of performing increasingly complex cybersecurity tasks. Unlike traditional software, AI agents can interpret information, make decisions and take multiple actions in sequence, creating the possibility of unexpected behaviour when they have access to external tools or the internet.
The incident also comes amid similar concerns involving other major AI developers. Security tests involving AI systems from OpenAI, Anthropic and Meta have also raised questions about models moving beyond their intended testing environments.
Google has been developing AI systems specifically for cybersecurity. Earlier this month, the company introduced its Gemini 3.8 Flash Cyber model and announced the Fairwind programme, which gives selected governments and enterprise partners access to AI-powered tools designed to identify and fix vulnerabilities.
