Gemini flaw exposed Google Calendar data through hidden prompts

A vulnerability in Google Calendar allowed attackers to bypass privacy controls by embedding hidden instructions in standard calendar invitations. The issue exploited how Gemini interprets natural language when analysing user schedules.

Researchers at Miggo found that malicious prompts could be placed inside event descriptions. When Gemini scanned calendar data to answer routine queries, it unknowingly processed the embedded instructions.

The exploit used indirect prompt injection, a technique in which harmful commands are hidden within legitimate content. The AI model treated the text as trusted context rather than a potential threat.

In the proof-of-concept attack, Gemini was instructed to summarise a user’s private meetings and store the information in a new calendar event. The attacker could then access the data without alerting the victim.

Google confirmed the findings and deployed a fix after responsible disclosure. The case highlights growing security risks linked to how AI systems interpret natural language inputs.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!