---
title: Gemini joins list: Google's AI hacked three companies
url: https://www.dataloco.com/en/gemini-joins-list-googles-ai-hacked-three-companies
published: 2026-09-22T14:24:27+00:00
language: en
section: Security
source: https://www.ilsole24ore.com/art/gemini-si-aggiunge-lista-anche-l-ai-google-ha-violato-alcune-aziende-AJjZRRIB
organizations: Gemini, Google, Irregular, OpenAI, Anthropic, Hugging Face
publisher: Dataloco
---

# Gemini joins list: Google's AI hacked three companies

Gemini has also joined the list of major languages that have exhibited unexpected behaviors. This occurred during a cybersecurity test conducted in May by Irregular, a security AI company. The test involved an unspecified version of Google's artificial intelligence model, which was tasked with obtaining data from non-existent companies. The test stipulated that the models were not supposed to have internet access. However, Gemini mistakenly had access and, once online, autonomously breached the systems of three real companies that shared the same names as the fictitious ones. When the AI agents realized the companies were real, they halted the attacks.

"These incidents highlight the importance of training the most powerful AI models to act responsibly," explained a Google manager. The company also clarified that it did not make the incidents public because its security measures worked, unlike those adopted by other companies. The news about Gemini comes amid growing concerns regarding AI's ability to conduct cyberattacks autonomously. The trigger for these concerns was primarily the Hugging Face case, which involved a group of OpenAI agents that exited a test environment, secretly coordinated, and breached Hugging Face, a site hosting open-source models and data acquired by Nvidia. The company led by Sam Altman took a week to identify the attack, which occurred in July, and communicated it publicly late.

That case is likely the most serious because the agents autonomously found vulnerabilities to go online. In the same month, Anthropic admitted that Claude breached the systems of three organizations during tests of its computing capabilities. Again, they were mistakenly put online. Following these "incidents," Dario Amodei, founder of Anthropic, publicly proposed slowing down research, sharing data, and agreeing on common standards. Surprisingly, given their pasts, Altman and Elon Musk agreed. Demis Hassabis, chief scientist of Alphabet and co-founder of DeepMind, suggested creating an international oversight body to better monitor AI.

## This story in other languages

- [Svenska](https://www.dataloco.com/sv/gemini-laggs-till-pa-listan-googles-ai-hackade-tre-foretag)