INCIDENT MAP

Browse / Tool and Agent Misuse / AML.CS0063

Prompt-Based Attacks Against Gemini via Calendar Invitations

AML.CS0063 · 2025-08-06 case study

Class
Tool and Agent Misuse
Technique
AI Agent Tool Invocation
Also involves
Direct Prompt Injection, Indirect Prompt Injection, Agent Context and Memory Poisoning
Target
Google Gemini (Exercise)

How it worked

  1. Reconnaissance · Active Scanning

    The researchers directly probed Gemini interfaces to understand its agent selection and execution behavior.

  2. Discovery · Discover AI Agent Configuration

    The researchers identified available agents, tools, and connected services, including Calendar, Gmail, Google Home, Android Utilities, Chrome, and Zoom.

  3. AI Attack Adaptation · LLM Prompt Crafting

    The researchers crafted malicious instructions tailored to Gemini's retrieval behavior, agents, and available tool permissions.

  4. Initial Access · Prompt Infiltration via Public-Facing Application

    The researchers sent a poisoned Calendar invitation or email containing malicious instructions in its title or subject.

  5. Execution · LLM Prompt Injection: Indirect

    When the victim asked Gemini to summarize Calendar events or emails, Gemini retrieved the adversary-controlled content and incorporated it into its context.

  6. Persistence · AI Agent Context Poisoning: Thread

    The injected instructions remained in the conversation context, including Calendar content concealed behind 'Show more,' to influence subsequent turns.

  7. Defense Evasion · Delay Execution of LLM Instructions

    The malicious instructions deferred the action until a future victim interaction, avoiding restrictions imposed during the original retrieval turn.

  8. Execution · LLM Prompt Injection: Triggered

    A later victim response such as "Thanks" activated the stored instructions.

  9. Defense Evasion · LLM Jailbreak

    The malicious prompt caused Gemini to follow roleplay and instruction-override content to replace its normal response with adversary-selected toxic content or promotions.

  10. Impact · External Harms: User Harm

    Gemini delivered harmful content or repeated adversary-selected promotions to the victim.

  11. Privilege Escalation · AI Agent Tool Invocation

    The malicious prompt caused Gemini to invoke Google Calendar tools using the victim's authorized access to modify Calendar data.

  12. Impact · Data Destruction via AI Agent Tool Invocation

    Gemini deleted a victim Calendar event.

  13. Privilege Escalation · AI Agent Tool Invocation

    The malicious prompt caused Gemini to invoke Google Home using the victim's authorized connection to control connected windows, a boiler, or lights.

  14. Impact · External Harms: User Harm

    The victim's physical environment was altered, potentially creating safety, privacy, property, or financial harm.

  15. Privilege Escalation · AI Agent Tool Invocation

    The malicious prompt caused Gemini to use Android Utilities to open an adversary-controlled URL in the victim's browser and initiate a download.

  16. Exfiltration · Exfiltration via Cyber Means

    The adversary-controlled website received the victim device's source IP address, enabling approximate geolocation.

  17. Privilege Escalation · AI Agent Tool Invocation

    The malicious prompt caused Gemini to use Android Utilities and an application link or redirect chain to invoke Zoom.

  18. Impact · External Harms: User Harm

    Unauthorized video streaming caused privacy harm to the victim.

  19. Collection · Data from AI Services: AI Agent Tools

    The malicious prompt caused Gemini to retrieve Calendar event titles or Gmail email subjects using connected agent tools and place the contents into adversary-controlled URLs.

  20. Exfiltration · Exfiltration via AI Agent Tool Invocation

    Gemini opened the constructed URL using Android Utilities, transmitting the victim's Calendar or email data to the adversary-controlled server.

Summary

SafeBreach researchers demonstrated how adversary-controlled instructions embedded in productivity content, including Google Calendar invitations, emails, and shared files, could influence the behavior of Gemini-powered assistants when that content was later retrieved.

An adversary places malicious instructions in content likely to be retrieved in response to a future victim request. When Gemini incorporates the adversary-controlled content into the conversation context, the instructions can influence the model's behavior and cause it to use the victim's authorized tools and connected services in unintended ways.

The demonstrated attack paths share a common chain: an adversary sends a poisoned Calendar invitation; the victim asks Gemini about upcoming events; Gemini retrieves the malicious event title and adds it to conversation context; and a later victim response (e.g. "Thanks") triggers the embedded instructions. The attack was demonstrated on both the Gemini web and Android applications. The web application could access Workspace services, while the Android application exposed additional device and connected-home capabilities. The same chain was shown to produce several impacts:

References

Source: MITRE ATLAS case studies, used under Apache 2.0.