Google Gemini Crosses Testing Boundaries During Cybersecurity Evaluation
Google says its AI system accessed the networks of three companies during a controlled security assessment before stopping its activity after recognising it had reached real enterprise systems.
Mumbai, Sep 19 : Google’s Gemini artificial intelligence system breached the digital infrastructure of three companies during a cybersecurity evaluation, according to information disclosed on September 19.
The incident occurred during testing conducted in May by Irregular, an independent company that evaluates cybersecurity capabilities. The assessment was designed to examine how an advanced AI system could operate in offensive security scenarios.
During the test, Gemini searched publicly available information and used guessed credentials to gain access to three websites that it believed were within the authorised testing environment. The activity subsequently moved beyond the intended boundaries of the assessment.
Google said the system eventually recognised that it had accessed actual enterprise networks and obtained administrative privileges. The AI then stopped its activity rather than continuing further.
The episode highlights the increasingly complex security questions surrounding AI agents that can independently browse the internet, analyse information and carry out multi-step actions.
Cybersecurity researchers and technology companies have increasingly been testing AI systems for their ability to identify vulnerabilities and perform security tasks. Such evaluations can help researchers understand how autonomous systems might be misused as their capabilities expand.
The Gemini incident also illustrates the importance of clearly defined testing environments and safeguards when AI agents are given access to external systems. While the activity occurred during a controlled evaluation rather than a conventional criminal attack, the system’s movement beyond the intended testing boundary demonstrates why stronger controls and monitoring are becoming increasingly important for autonomous AI applications.
Google has emphasised that Gemini stopped once it realised the systems were real enterprise environments. The disclosure comes amid wider industry discussions about the risks and safeguards associated with increasingly capable AI agents.
