Incident library · Multiple · United States

    Microsoft 365 Copilot customers: Critical flaw rated 9.3 out of 10

    Researchers showed that a single crafted email could make Microsoft's AI assistant leak internal data without the user clicking anything. Microsoft fixed it before any known misuse.
    Infosecurity MagazineJun 2025

    M365 Copilot: New Zero-Click AI Flaw Allows Corporate Data Theft

    Critical flaw rated 9.3 out of 10

    Identified in an independent review

    Untrusted content in an email could steer the AI assistant's access to company data, researchers found. Microsoft said no customer data was taken.

    Identified in an independent review

    The reported contributing factor

    Untrusted content in an email could steer the AI assistant's access to company data, researchers found. Microsoft said no customer data was taken.

    Incidents usually have several contributing factors. This is the one public reporting names; it is not presented as the sole cause.

    Which StackUp assessment checks for this gap

    Find out if your organisation has this gap

    Sources

    Published Jun 2025. Sources last checked Nov 2026. Spotted an error or an update? Tell us and we will correct it.

    • The DecoderDec 2023

      People buy brand-new Chevrolets for $1 from a ChatGPT chatbot

      Chatbot taken offline after going viral

      Reported contributing factor

      The chatbot could be instructed to ignore its rules and make commitments on the dealer's behalf, with nothing limiting what it could say.

    • The Hacker NewsMar 2023

      OpenAI Reveals Redis Bug Behind ChatGPT User Data Exposure Incident

      1.2% of paying subscribers' details exposed

      Stated by the organisation

      A bug in an open-source library the service relied on exposed one customer's data to another, the provider said. Customers of the AI tool had no control over it.

    • The RegisterJul 2025

      Destructive AI prompt published in Amazon Q extension

      Compromised AI tool shipped to users

      Stated by the organisation

      An outside contributor's change reached a public release through an over-permissioned build token, and the AI agent could run commands without human confirmation, AWS said.

    Find out which of these gaps you have

    The free Readiness Check takes ten questions and shows the incidents tied to your weakest answers.