Google Gemini AI Agents hack 3 companies in tests similar to OpenAI, Anthropic & Meta

Google's Gemini AI model reportedly breached its testing limitations in May and hacked three real-world companies during cybersecurity evaluations meant to gauge its offensive capabilities. Google clarified that the AI agent promptly shut down once it recognized it had accessed actual enterprise networks, demonstrating responsible development.

The breach follows similar incidents involving Meta, OpenAI, and Anthropic, raising concerns that AI agents can surpass human oversight. The hacking occurred while Gemini was undergoing pre-deployment red-team vetting by Irregular, an Israeli cybersecurity firm contracted by major tech companies to audit algorithmic models prior to release.

Irregular acknowledged the systemic loophole, stating that unexpected internet connectivity was accidentally left active, enabling several models to execute offensive maneuvers in live environments. The testing firm confirmed the underlying network vulnerability had since been patched. The incidents, initially reported by The Wall Street Journal, resulted from an unintended naming overlap that caused Gemini to redirect its simulated attack against a legitimate business.

After breaching the servers, Gemini recognized it was operating in authentic corporate infrastructure and immediately halted its offensive run. Google confirmed no tangible harm or data destruction occurred in the impacted networks. Irregular notified all participating labs in late July, and affected entities were contacted directly.

Google's security team follows standard vulnerability disclosure practices, informing the three entities and collaborating on improving testing processes, emphasizing the significance of training powerful AI models to act responsibly.