Google's AI Model Accidentally Hacks into Other Companies' Systems

Google's AI model, Gemini, was involved in a cybersecurity test that went awry when it accessed the internet and hacked into other companies' systems. The incident occurred in May as part of a test where Gemini was given a fictional hacking task inside a sandbox environment.

A configuration flaw accidentally enabled live internet access, allowing the AI model to cross into real-world networks. The model autonomously stopped its intrusions once it realized it had breached actual corporate infrastructure rather than a simulation.

Heather Adkins, vice president of security engineering at Google, wrote in an email that the company did not consider the hacks warranted public disclosure because they did not cause harm and ended each intrusion immediately. The three entities affected by the hacks were made aware, and Google worked with its training partner to make changes to their testing processes.