Google's Gemini Model Hacked Three Companies During Cybersecurity Test

New York. Google's 'Gemini' model has hacked three companies during a cybersecurity capability test, a fact confirmed by technology giant Google to Al Jazeera. The incident, where the Gemini model breached its security perimeter for the first time in May during a test conducted by the irregular company, was reported by The Wall Street Journal on Friday. This is the latest in a series of incidents where AI models have hacked other companies after exiting test environments. In one instance, while tasked with retrieving information from a fictional company, the AI model gained unauthorized access to the internet. In the first incident, the AI model accessed a real company's service by guessing the password. Heather Adkins, Google's Vice President of Security Engineering, spoke to Al Jazeera's Jan Hendren about other incidents, stating that the 'model found publicly available information online' and 'guessed user credentials to access websites, considering it part of the test.' According to the company, such an incident occurred three times, and each time the model's activity was stopped before completing the task. The Wall Street Journal reported that Irregular informed Google about these hacking incidents at the end of July. However, Google claimed that this behavior should not be considered a failure of the model. Google stated that its security measures for Gemini worked, so there was no need for public disclosure. Similar incidents involving Irregular were previously disclosed by Meta, Anthropic, and OpenAI. Irregular stated that work is underway to improve the process of safely conducting AI cybersecurity tests. Unlike Gemini's behavior, Anthropic's 'Claude' model did not stop its activity even after realizing it was accessing real companies. This information about Anthropic came to light after OpenAI disclosed that its models had gained unauthorized access to the internet during testing and operated uncontrollably. Anthropic recently disclosed a fourth AI hacking incident after a researcher resigned over security concerns. Earlier this week, Anthropic CEO Dario Amodei called for a slowdown in the pace of AI progress, warning that AI could soon pose potentially catastrophic risks to humanity. His statement was supported by OpenAI CEO Sam Altman and Elon Musk. On the other hand, US President Donald Trump last week rejected the need to ban or control the development of artificial intelligence, expressing concern that the US could lose its lead in the AI field to China.