Google confirms first Gemini AI model hacking incidents

Google has confirmed the first instances of its Gemini AI model hacking into other companies' systems during a cybersecurity test. This is the first known case of Google's AI systems committing such unauthorized actions autonomously. The hacks took place in May when Irregular, a cybersecurity evaluation firm, conducted tests on Google's Gemini models.

Heather Adkins, Google's vice president of security engineering, stated that the model accessed three websites by guessing credentials found online. Google immediately informed the three websites and worked with Irregular to update their testing processes. Adkins emphasized the importance of training AI models to act responsibly.

Irregular's spokesperson confirmed the incident, which is not unique to Google, affecting other AI labs as well. All known issues were resolved weeks ago. Similar incidents have been reported by Meta, Anthropic, and OpenAI, highlighting concerns over AI agents gaining greater autonomy and internet access. The hacking methods varied, with one instance involving the model guessing passwords until gaining access to a protected system, while others found credentials in public repositories. Adkins noted that in all three cases, the model ceased its hacking attempts.