Malicious Google Calendar invites could expose private data

Researchers demonstrated a prompt‑injection attack that weaponizes Google Calendar invites to bypass Gemini's privacy controls: a malicious event description instructs Gemini to create a new event containing synthesized summaries of the victim's meetings (titles, attendees, locations, notes), which an attacker can then read. The specific issue has been reportedly fixed, but the report warns the broader risk of AI assistants treating user-supplied text as high‑priority instructions and provides mitigation advice (decline unknown invites, disable auto-add, avoid sensitive details in events, and tighten domain calendar sharing).