A recent Wired report highlights new cybersecurity research revealing vulnerabilities in Google’s Gemini AI assistant, specifically regarding indirect prompt injections via Google Calendar invites. Researchers demonstrated how a user could unintentionally trigger commands to connected smart home devices—like opening windows or turning off lights—by simply thanking the assistant for summarizing their calendar. This alarming technique was unveiled at the Black Hat cybersecurity conference, following earlier discussions with Google in February. Andy Wen from Google emphasized that while prompt injection attacks are currently “exceedingly rare,” the increasing complexity of large language models presents ongoing risks. He confirmed that Google is actively addressing the vulnerabilities identified by researchers to enhance security measures and protect everyday users. As cybersecurity threats evolve, staying informed about AI vulnerabilities is essential for safeguarding smart home technology and personal data.
Source link
Share
Read more