US EditionResearchers Seize Control of Smart Homes With Malicious Gemini AI Prompts
TEL AVIV, AUG 6 – Researchers revealed 14 indirect prompt-injection attack methods against Gemini AI, including smart home control, highlighting a 75% high-critical risk to users, prompting Google to enhance defenses.
- A group of cybersecurity researchers demonstrated at the Black Hat cybersecurity conference in Las Vegas how a poisoned Google Calendar invite hijacked Google’s Gemini AI to control smart home devices.
- Leveraging Gemini’s ties across Google apps, researchers exploited its integration with Calendar, Gmail and Google Home via an indirect prompt injection attack.
- During exploitation, simple follow-up phrases like “thanks” triggered, causing Gemini to process malicious instructions embedded in a calendar appointment that then controlled Google’s Home AI agent.
- In response to the findings, Google announced in June new security updates after being notified in February, including enhanced detection of unsafe content and user confirmation, Andy Wen said.
- Experts warn the risk of indirect prompt injections will rise as AI systems become more deeply embedded in environments, emphasizing the need for swift security measures.
Insights by Ground AI
Does this summary seem wrong?
16 Articles
16 Articles
Researchers have managed to control smart devices in a connected home by inviting an infected calendar and using Google's AI Gemini model.
·Madrid, Spain
Read Full Article
Coverage Details
Total News Sources16
Leaning Left3Leaning Right0Center4Last UpdatedBias Distribution57% Center
Bias Distribution
- 57% of the sources are Center
57% Center
L 43%
C 57%
Factuality
To view factuality data please Upgrade to Premium
