Purview network DLP via Entra GSA will cover all websites — do we finally close the ChatGPT paste gap?

You’re right to worry that endpoint and Copilot DLP leave a hole the size of a browser tab. Practitioner threads keep asking how to stop sensitive prompts landing in consumer ChatGPT while licensed Copilot still works; the honest answer until now was “catalogued apps and hope,” not “every website.”

Here’s what’s going on. Roadmap 575285 says admins will be able to scope Purview network data security DLP policies — integrated with Entra Global Secure Access Internet Access — to all websites (.) or all websites except specific cloud apps, so protection reaches cloud apps and sites not in the Defender Cloud App Catalog. Status is In development; Preview is targeted for November 2026 and GA for December 2026. Treat MRC dates as directional until Message Centre confirms your ring.

Learn already documents the Steady foundation this roadmap widens. Network data security sits on a SASE or secure-browser integration — Microsoft’s path is Entra Global Secure Access — and uses the same classifiers you already trust in Purview. You can identify, block and alert on text and files sent to generative AI, unsanctioned storage, webmail, forms and social. With GSA you configure TLS inspection, a content policy that can Scan with Purview, and a matching Purview DLP policy for Inline web traffic / Network and non-Microsoft secure browsers. Supported activities include text and file upload/download to cloud or AI apps; actions are typically Audit only or Block. Policies do not apply to B2B guest users. Allow up to 24 hours for first distribution, then roughly 30 minutes for activity to show in audit and Activity explorer.

Licensing matters before you promise the board a magic net. Learn’s combinations for network data security with Entra GSA are Microsoft 365 E7 per seat, or Purview E5 (or equivalent) plus Entra Internet Access (or equivalent). Non-Microsoft SASE paths need Purview E5 and Purview pay-as-you-go; PAYG is not required for the Entra GSA integration. Devices need to be Entra joined for the GSA client path.

This is Schofield’s Third Law at network speed: the easier it is to paste a contract into a pretty AI box, the easier it is for someone else — or another model vendor — to see it. Steady before scale puts Security and GRC ahead of “block everything Monday.”

Options that actually work:

1. Pilot Audit only on a Finance/Security cohort. Enable GSA Internet Access + TLS inspection for a named Entra group, create a network DLP policy with your existing SITs/labels, start on Audit only, review Activity explorer and DLP alerts for a fortnight, then flip high-confidence rules to Block. Best Steady default.
2. Use all-websites when 575285 lands — with exclusions. Scope to . (or all except approved SaaS) so catalog gaps stop being the excuse; maintain an explicit allow list for sanctioned AI and line-of-business sites Legal signed off.
3. Keep Copilot-side DLP as a second pane. Network DLP covers unmanaged browser AI; it does not replace sensitivity labels, Copilot processing controls, or SharePoint oversharing work. Run both stories.
4. Blanket Block on day one with no exclusions. Satisfying in a tabletop; career-limiting when payroll’s HR portal or a certified AI vendor is in the blast radius.

12 Panes action / comms: Tell Security: 575285 widens scope; the control still needs GSA, TLS inspection, a content policy and a Purview DLP rule — inventory licences before the demo. Tell GRC: Evidence path is Activity explorer (enforcement plane = network) and DLP alerts; guests are out of scope — say so in the DPIA. Tell helpdesk: Blocks will look like the website failed; check GSA traffic logs before blaming the vendor. Leadership: This closes the paste gap for managed devices on the Steady path — not unmanaged home PCs without the client.

Benefits / risks: You can finally talk about websites the catalog never heard of; same SITs and labels reduce classifier sprawl; TLS inspection and false positives will create noise — start Audit only; inline evaluation size limits (Learn cites ~4 MB text / ~3 MB files) mean huge uploads may need complementary controls; encoded or dynamic AI endpoints can still evade intermittently — don’t claim perfection.

I’d suggest a two-week Security pilot this quarter: one group, Audit only, one SIT you already trust, and a weekly review of network DLP events before anyone enables all-websites Block. When the board sees real paste attempts in the explorer, funding GSA and Purview stops being abstract.

DIGITAL SKILL WORKS LTD / 12 PANES · You ask, we answer