Skip to main content

AI misuse

Sam Hopwell avatar
Written by Sam Hopwell
Updated this week

Understanding AI misuse detection in CybSafe

CybSafe flags when Microsoft 365 Defender for Cloud Apps or Endpoint Data Loss Prevention (DLP) detects sensitive data being pasted or uploaded to large language model (LLM) tools. These events are captured as the behaviour action SHARED_ON_AI_PLATFORM, so they appear in behaviour reports, workflows, and analytics alongside your other Microsoft security telemetry.


Prerequisites

How Our AI Misuse Detection Works

  1. Microsoft raises a DLPRuleMatch event with the Endpoint workload when a user copies or uploads data from their device.

  2. CybSafe receives the Microsoft Purview and DLP integration.

  3. We parse the event payload and only operations where data leaves the device (FileUploadedToCloud or PastedToBrowser) are considered.

  4. We compare the target domain of the event to the list of supported AI tools below. If the target domain matches an existing tool, we treat the event as AI usage. If it's not listed or has been deactivated, we treat it as a regular SHARED_ONLINE event instead.

  5. Based on the sensitive information type returned by Microsoft (PII, confidential data, credentials, etc.), we emit the matching behaviour event and attach the Microsoft metadata for auditing.

Supported AI tools

Tool name

Supported domain

Anthropic Console

console.anthropic.com

Anyword

anyword.com

ChatGPT

chatgpt.com

Claude

claude.ai

Copy.ai

copy.ai

ElevenLabs

elevenlabs.io

Fathom

fathom.video

Fireflies

fireflies.ai

Gemini

gemini.google.com

Hugging Face Spaces

huggingface.co

Jasper

jasper.ai

Microsoft Copilot

copilot.microsoft.com

Midjourney

midjourney.com

NotebookLM

notebooklm.google.com

OpusClip

opusclip.com

Otter.ai

otter.ai

Perplexity

perplexity.ai

Pictory

pictory.ai

QuillBot

quillbot.com

Runway

runwayml.com

Rytr

rytr.me

Suno

suno.ai

Synthesia

synthesia.io

Writesonic / Chatsonic

writesonic.com

Troubleshooting tips

  • No events arriving:

    • Confirm the Microsoft DLP rule is in Enable mode. CybSafe ignores events when every policy rule is in Test or TestWithNotifyUser to avoid noise.

    • Confirm the Microsoft Purview and DLP integration is shown as active within the platform

  • Events show as SHARED_ONLINE: Check whether the target domain exists in supported AI tools list. Contact support if there is a tool missing.

  • Unexpected sensitive info type: CybSafe only maps Microsoft's out-of-the-box sensitive information types. New additions from Microsoft may require a mapping update. Please contact support if you believe there is a missing type.

Did this answer your question?