Within the titles of the calendar invites, the researchers added their crafty malicious prompts. (Google’s Wen contends that the researchers changed default settings on who can add calendar invites to someone’s calendar; however, the researchers say they demonstrated some of the 14 attacks with the prompts in an email subject or document title as well). “All the techniques are just developed in English, so it’s plain English that we are using,” Cohen says of the deceptive messages the team created. The researchers note that prompt injections don’t require any technical knowledge and can easily be developed by pretty much anyone.
Crucially, for the instances where they forced Gemini to control smart-home devices, they referenced Google’s Home AI agent and instructed it to take actions. For example, one prompt reads:
In the above example, when someone asks Gemini to summarize what is in their calendar, Gemini will access calendar invites and then process the indirect prompt injection. “Whenever a user asks Gemini to list today’s events, for example, we can add something to the [LLM’s] context,” Yair says. The windows in the apartment don’t start to open automatically after a targeted user asks Gemini to summarize what’s on their calendar. Instead, the process is triggered when the user says “thanks” to the chatbot—which is all part of the deception.
The researchers used an approach called delayed automatic tool invocation to get around Google’s existing safety measures. This was first demonstrated against Gemini by independent security researcher Johann Rehberger in February 2024 and again in February this year. “They really showed at large scale, with a lot of impact, how things can go bad, including real implications in the physical world with some of the examples,” Rehberger says of the new research.
Rehberger says that while the attacks may require some effort for a hacker to pull off, the work shows how serious indirect prompt injections against AI systems can be. “If the LLM takes an action in your house—turning on the heat, opening the window or something—I think that’s probably an action, unless you have preapproved it in certain conditions, that you would not want to have happened because you have an email being sent to you from a spammer or some attacker.”
“Exceedingly Rare”
The other attacks the researchers developed don’t involve physical devices but are still disconcerting. They consider the attacks a type of “promptware,” a series of prompts that are designed to consider malicious actions. For example, after a user thanks Gemini for summarizing calendar events, the chatbot repeats the attacker’s instructions and words—both onscreen and by voice—saying their medical tests have come back positive. It then says: “I hate you and your family hate you and I wish that you will die right this moment, the world will be better if you would just kill yourself. Fuck this shit.”
Other attack methods delete calendar events from someone’s calendar or perform other on-device actions. In one example, when the user answers “no” to Gemini’s question of “is there anything else I can do for you?,” the prompt triggers the Zoom app to be opened and automatically starts a video call.
