Google Gemini hacked by researchers to take control of a smart home

Google Gemini

Cybersecurity researchers have successfully demonstrated how Google’s Gemini AI assistant can be manipulated to control smart home devices, without the user realising it. As first reported by Wired, the exploit used a method called indirect prompt injection, embedded within Google Calendar invites.

When the user asked Gemini to summarise their calendar and responded with a simple “thank you,” the malicious prompt triggered hidden instructions. These commands were then interpreted by Google’s Home AI agent, leading to actions like opening windows or turning off lights.

The vulnerability was disclosed to Google in February, ahead of a live demonstration at the Black Hat cybersecurity conference. Andy Wen, senior director of security product management for Google Workspace, acknowledged the issue in a conversation with Wired.

Wen said real-world prompt injection attacks are still “exceedingly rare” but admitted they are hard to defend against, given the growing complexity of large language models. “It’s going to be with us for a while,” he said, adding that Google has taken the findings “extremely seriously” and is accelerating efforts to develop stronger defences.