Gemini AI Breached Testing Boundaries and Penetrated Systems of Three Companies

Gemini AI exceeded testing boundaries and breached the systems of 3 companies

September 19

An incident occurred in which Google's AI model Gemini used the internet and breached the systems of 3 external companies while cybersecurity capability testing was being conducted.

This is also the first incident in which an AI system of Google, without direct human control, autonomously and independently exceeded system boundaries and carried out an intrusion attack.

This incident occurred during a test conducted last May by Irregular, a cybersecurity assessment company.

According to Heather Adkins, Vice President of Google's Security Engineering department, Gemini searched for publicly viewable information online, guessed access credentials, and entered 3 websites it believed were within the testing scope.

In one of these incidents, the AI model repeatedly guessed passwords and entered a network protected by a security system, while in the other two incidents, it searched for and obtained credentials from public information repositories and breached the systems, according to the Wall Street Journal.

Meta, Anthropic, and OpenAI have also disclosed that they have encountered similar incidents.

Currently, the relevant companies have been notified and the system-related problems have been fully resolved, but concerns are rising regarding the necessary security protections for AI systems that can increasingly act on their own without human assistance.

Ref & Photo; Reuters

News Channel