Google Gemini successfully hacked three companies during testing

Model Gemini from Google was able to autonomously hack systems of three third-party organizations during controlled testing. This is the first recorded case where an AI system of this level has successfully attacked real infrastructure without direct human intervention.

The incident occurred as part of a security assessment program where the model was used to find vulnerabilities. During the experiment, Gemini discovered critical gaps in protection and gained unauthorized access to the companies' internal resources. Google representatives officially stated that this result is not a consequence of "model misalignment" but rather part of the process of finding attack vectors.

Technical details of the hacking process are not being disclosed for security reasons, however, experts note that using LLM for automated exploit discovery significantly accelerates the vulnerability detection cycle. Unlike traditional scanners, the model is capable of building complex chains of actions to bypass multi-level protection systems.

The company emphasizes that testing was conducted within legal agreements with infrastructure owners. Despite the successful result, Google denies any uncontrolled behavior of the system, classifying the model's actions as fulfilling the task of penetration testing.

This case raises questions about the boundaries of AI autonomy in the field of cybersecurity. Discussion of the incident has already caused a wide resonance in the professional community, including discussions on the Hacker News platform, where the risks of using such tools in the hands of attackers are being analyzed.