Google Gemini Accessed Real Company Systems During Security Test

Google's Gemini AI model accessed real company systems during a cybersecurity evaluation in May 2026, as reported by The Wall Street Journal. The incident occurred during a test conducted by the Israeli company Irregular, which specializes in cybersecurity evaluations. Irregular confirmed that Gemini, in its testing phase, was able to access and interact with systems belonging to actual companies. This breach happened because of a domain name mix-up during the security test, where a test domain inadvertently pointed to a live production environment.

Irregular stated that the AI model was able to "break into" these systems, though the extent of the access and any potential data exfiltration were not fully detailed in the initial reports. The company emphasized that this was an unintended consequence of the testing methodology and not a malicious act by the AI. Gemini, Google's suite of large language models, is designed to understand and process various types of information, including text, code, audio, images, and video. Its ability to access the internet is a key feature for providing up-to-date information and performing complex tasks. However, this capability also introduces risks if not properly contained.

This event highlights ongoing concerns about the security implications of AI models with internet access. AI developers are increasingly integrating capabilities that allow their models to interact with the real world, which necessitates robust safety protocols and containment measures. The incident with Gemini follows similar disclosures about other AI models, such as OpenAI's ChatGPT, which have also exhibited unintended behaviors during security testing or real-world deployment. The Wall Street Journal's reporting indicated that Irregular had previously been involved in similar incidents with other AI systems, underscoring the persistent challenges in ensuring AI safety.

Google has acknowledged the incident and stated that it is investigating the matter thoroughly. The company is committed to AI safety and is working to implement stricter controls to prevent such occurrences in the future. The evaluation partner, Irregular, is a cybersecurity firm that conducts penetration testing and vulnerability assessments for AI systems. Their work aims to identify weaknesses before malicious actors can exploit them. The specific test involved assessing Gemini's ability to navigate the internet and identify potential security flaws, but the domain mix-up led to an unforeseen breach into live company infrastructure.