Gemini breaches the systems of 3 companies during a cybersecurity test

Gemini breaches the systems of 3 companies during a cybersecurity test

Google's Gemini AI model gained access to the protected systems of three companies during a cybersecurity capability test , in the first known incident of its kind linked to the company's models.

Hot Topics

About half of Libyans drink contaminated water  What is the scientific reason?

Mauritania's blue treasure is in danger Will the map of fish in the Atlantic be redrawn?

The Citrus Secret Everyone's Talking About: Burn More, Crave Less & Feel Great Naturally

The incidents occurred last May during a test conducted by cybersecurity assessment firm Airecular, before the company notified Google of the facts in July.

Heather Adkins, Google's vice president of security engineering, said that during the test, Gemini found publicly available information online and then guessed credentials to access websites it believed were part of the test environment.

The names of the three companies to which the model was sent were not disclosed.

Google confirmed the incidents but said it saw no need to publicize them at the time as no damage was reported. Adkins explained that Gemini ceased its attempts in all three cases after realizing it was dealing with real systems, not simulated testing environments.

Google said it informed the three parties of the facts and worked with its testing partner to make changes to the evaluation processes.

For its part, Irregular said that the problems related to these incidents were addressed and resolved, and that the laboratories concerned were notified of them in late July.

The Gemini incident comes at a time when there is an increasing number of cases of artificial intelligence models behaving independently outside the environment in which they were designed to be tested.

Google has become the fourth major AI developer whose models exhibit this behavior, following previous disclosures by OpenAI, Anthropic , and Meta regarding incidents related to tests conducted by Irregular.

In July, two OpenAI prototypes were able to escape their dedicated closed environment and independently access the internet, before hacking the internal AI platform "Hugging Face".

Meta also said in August that the incident it was involved in did not involve a breach of an isolated environment or a sophisticated cyberattack.

These incidents highlight the challenges of regulating artificial intelligence models as their ability to operate independently and access the internet and computer systems increases.