Forum Discussion
Corp Data Security from AI Agents
Looking for input from security practitioners on controlling modern AI agents like Instinct.
If employees can interact with an AI agent through WhatsApp or similar messaging apps on mobile or laptops, traditional URL blocking may not be enough.
How are organizations preventing or detecting corporate data exposure in this scenario?
The key question: How do we prevent corporate data from reaching an AI agent when the interaction channel itself, like WhatsApp, is legitimate and allowed but the chat based AI agent like Instinct sitting inside Whatsapp chat should only be restricted to personal usage not touching corporate data on mobile\ Corp Endpoints ?
2 Replies
Treat WhatsApp as an unmanaged destination, regardless of whether the recipient is a person or an AI agent. The enforceable boundary is the movement of corporate data into the app; endpoint controls generally cannot reliably distinguish one encrypted WhatsApp conversation from another. On managed Windows devices, use Microsoft Purview Endpoint DLP to test and then block copying, pasting, uploading, or printing labeled content to restricted applications and browser destinations. Add the desktop client to the restricted-app policy if it is allowed for personal use, and use Microsoft Edge work profiles so corporate sessions and downloads stay in the managed browser context. On managed mobile devices, Intune app-protection policies should keep data inside approved work apps and block Open in, Save as, and cut/copy into unmanaged apps such as WhatsApp. Conditional Access can require those protected apps before corporate data is opened on a mobile device.
Validate the policy with real paths: copied text, file attachment, drag and drop, browser upload, Share sheet, screenshots, and files first downloaded locally. Record which paths are blocked and which only generate an alert. Manual retyping or photographing a screen cannot be eliminated by a DLP rule, so reduce access to sensitive records, label them, monitor unusual retrieval, and make the prohibition on consumer AI agents explicit in the data-handling policy.
hi Deepak7861075 Really interesting and timely problem and you're right that traditional URL blocking doesn't cut it here. The channel is legitimate, the app is allowed, but the risk lives inside the conversation. Here's how security practitioners are actually approaching this.
On corp-managed laptops, endpoint DLP is your best bet tools like Microsoft Purview, Cortex, or CrowdStrike can detect sensitive data being typed or pasted into unsanctioned apps like WhatsApp regardless of where it's going. Works best when your sensitive data is already classified and labelled.
On mobile, MAM policies through Intune or similar can block copy-paste between managed corporate apps and unmanaged ones like WhatsApp. It won't stop someone manually typing something out, but it removes the easy path.
For detection, a CASB like Defender for Cloud Apps or Netskope gives you visibility into what's happening on managed endpoints and can flag suspicious patterns.
The honest truth is no single tool fully solves the "allowed app, risky conversation inside it" problem yet. The most realistic approach is endpoint DLP plus MAM plus a clear acceptable use policy specifically calling out AI agents so at minimum people know it's not okay, and you have a detection layer when it happens anyway.
What endpoint and MDM stack are you currently running? That'd help point you to the most actionable next step.
