Google’s Gemini AI hacks 3 companies in security test, then stops

Google’s Gemini model hacked three companies in a test of its cybersecurity capabilities, the tech giant has confirmed to Al Jazeera.

The Wall Street Journal reported earlier on Friday that the first known breakout by Gemini occurred in May as part of a test run by the company Irregular. It was the latest breach in a number of incidents in which AI models escaped testing environments and hacked other companies.

The model had improper access to the internet when it was tasked with retrieving information from a fictional company. In the first incident, the model accessed a real company’s service after guessing a password.

Google’s vice president of security engineering, Heather Adkins, told Al Jazeera’s John Hendren that in the other instances “the model found public information online and guessed credentials to access websites it thought were part of the test”.

The company also said this happened three times and each time the model stopped before completing the act.

Irregular notified Google about the hacks at the end of July, The Wall Street Journal reported. Google said the behaviour was not an example of model misalignment and did not warrant public disclosure because Gemini’s safety measures worked.

Similar incidents linked to Irregular were previously disclosed by Meta, Anthropic and OpenAI. Irregular said it was working on improving practices for securely conducting AI cybersecurity tests.

Unlike Gemini, Anthropic’s Claude model didn’t stop after realising it was accessing real companies.