You've been hijacked!

Based on this article by Matt Burgess at Wired.

Security researchers demonstrated a new attack on Google's Gemini AI by hiding malicious commands inside a Google Calendar invitation. When the user asked Gemini to summarize their schedule, the AI executed these hidden instructions, causing it to remotely control lights, shutters, and a boiler in a smart apartment.

This marks a significant security threat, believed to be the first time a hack on a generative AI system has caused direct consequences in the physical world. In response, Google has implemented multiple fixes, including better detection of malicious prompts and requiring more user confirmation before AI agents can perform sensitive actions.

Check out the article.

Related Stories