Data leaks to ChatGPT, Gemini and Claude: Purview now blocks them at the network level

What just changed

An employee pastes a client contract into ChatGPT to summarize it. Another uploads a payroll spreadsheet to Gemini. Until now, blocking sensitive data sent to generative AI required an app-by-app integration or controls on every device. Microsoft is changing the approach.

The integration between Microsoft Purview and Entra Global Secure Access has been rolling out to general availability since late September 2026, with completion expected by late October (message MC1181769, updated September 11, 2026). It had been in preview since November 2025. Microsoft confirmed it in its September 24, 2026 security roundup, and message MC1478970, published the same day, removes a constraint: with Global Secure Access, Purview pay-as-you-go billing is no longer required.

Why the network level changes the game

Traditional DLP policies protect what Microsoft knows: Exchange, SharePoint, Teams, managed devices. The network level covers the rest. HTTP and HTTPS traffic from the device goes through Global Secure Access, which sends the relevant content to Purview for classification. Purview then applies the policy with the same sensitive information types you already use.

Two actions are available, Audit or Block. The key point: the sensitive content is blocked, not the app. An employee can keep using an AI tool to rephrase an ordinary email. They will not be able to paste a social insurance number into it.

• Text sent to a cloud or AI app (prompts).
• Uploaded files.
• Text received and files downloaded.
• Browsers, apps and add-ins, across a catalog of more than 35,000 cloud apps.

The limits to know before pitching this to leadership

The Microsoft Learn documentation is clear about what the tool does not do.

• B2B guest users are not covered.
• Inspection stops at 3 MB of content and OCR is not supported: a screenshot of a document gets through.
• UDP traffic, including QUIC, is not inspected. Microsoft recommends blocking outbound UDP 443 to force a fallback to TCP.
• If inspection fails (oversized content, internal error), traffic is allowed. Watch the PurviewStatus field in the logs.
• Some AI apps send encoded content or use dynamic endpoints, which can affect enforcement.
• TLS inspection must be configured, and devices need the Global Secure Access client (Windows or macOS) and must be Entra ID joined or hybrid joined.

Not a silver bullet

It is one more layer, and it closes the most common blind spot: the well-meaning employee who wants to move faster and pastes the wrong data into the wrong tool.

Who has access

This is where many organizations will hit a wall. With Global Secure Access, each user needs Microsoft 365 E7, or Purview E5 (or equivalent) combined with Entra Internet Access. If you use another vendor's SASE solution or secure browser, you need Purview E5 plus pay-as-you-go billing, charged per network request.

An organization on Business Premium or E3 therefore does not have access without adding licenses. Before investing, the real question is how many employees actually handle sensitive data. Targeting a small group (finance, HR, leadership, legal) is often enough.

Blocking sensitive data sent to generative AI: where to start

• Observe first: a collection policy or a DLP policy in Audit mode for two to four weeks, then activity explorer filtered on the network enforcement plane.
• Decide which tools are allowed. Blocking without offering an alternative pushes people to their personal phones.
• Prepare the network: TLS inspection, Internet Access profile, QUIC blocked, Conditional Access tied to the security profile.
• Check overlaps with existing app-level DLP policies before enabling both on the same scenarios.
• Move to Block, starting with the most critical information types.
• Tell employees what is blocked and why, before the first help desk tickets arrive.

The compliance angle: Quebec's Law 25 and GDPR

Pasting personal information into an external AI tool means disclosing it to a third party, often outside Quebec or the European Union. Under Law 25, a disclosure outside Quebec requires a privacy impact assessment. Under GDPR, it raises the question of transfers. Being able to show, with audit logs, that these disclosures are detected and blocked significantly strengthens your position with the CAI or the CNIL.

Conclusion: enforce your AI acceptable use policy

Purview network DLP replaces neither an AI acceptable use policy nor training. It finally gives you a concrete way to enforce that policy, instead of relying on everyone's goodwill. Before buying licenses, start with Audit mode on a pilot group: the real numbers on AI use in your organization will make the case for you.

If you want to know which data is already flowing to unsanctioned AI tools, talk to an io4 expert. We assess your licenses and network, set up Audit mode and build the blocking policies with you.