Microsoft Brings Sensitive AI Upload Blocking to Entra Networks
Purview classifies files and text while Entra enforces network policy for employee and on-behalf-of agent traffic. Confirm forwarding and an active blocking rule before relying on it.
A sensitive company document can leave through an employee upload or an AI agent working on that employee’s behalf. Microsoft’s September 24 announcement says Purview and Entra Global Secure Access network data security is now generally available for both kinds of traffic. Purview classifies the file or text; Entra can enforce a policy at the network layer before an upload reaches an unsanctioned AI tool.
Microsoft describes an employee or on-behalf-of agent trying to upload a sensitive document and says “the policy can stop the transfer before the data leaves.” In your tenant, the upload must cross the network route and match an enforcing rule.
Where the network rule applies
Microsoft’s Global Secure Access documentation says Entra Internet Access acquires internet traffic through a user-aware forwarding profile, using either a desktop client or a remote network such as a branch. It lists network data loss prevention among the Internet Access features. The same documentation describes internet traffic logs that include enforced policy details. Start with the employee devices and agent egress locations you actually route through that profile.
Microsoft Purview’s DLP documentation describes collection policies for inline web traffic to unmanaged cloud apps. Purview evaluates content using sensitive information definitions and policy conditions; the enforcement action can block sharing. Its documented network option covers services including ChatGPT and Gemini. The policy needs to match the content, destination and activity you intend to control. Select inline web traffic for the outbound upload; SharePoint storage is a separate policy location.
This is a data path with two jobs: the traffic must pass through the inspection route, and the content policy must tell that route when to stop it. A browser session and an agent’s outbound request may originate in different places. Test both egress routes: the forwarding profile needs to carry each upload for this network rule to act.
Simulation shows matches without blocks
Purview’s deployment guidance recommends running DLP policies in simulation mode to tune scope and conditions before enabling enforcement. Defined actions are not applied in simulation mode. Use those matches to adjust sensitive-information definitions, covered users and destinations, then turn on the blocking rule when it fits the approved workflow. Microsoft says policies generally take effect about an hour after activation.
Purview sends monitored DLP activity to the Microsoft 365 Audit log and Activity explorer. Its documentation names filters for egress activity, matched policies and matched rules. Entra Internet Access supplies traffic logs with enforced policy details. Look at both views when a test misses: one tells you about content-policy matches, the other about the forwarded network traffic and its policy.
Test both upload routes
Confirm that the relevant internet forwarding profile covers the employee device and the agent’s egress location. Set a Purview inline-web collection policy for the sensitive content and risky destination in question; tune it in simulation, then enable a block for the approved scope. Use an authorized, nonproduction sample that matches the rule to test one human upload and one on-behalf-of agent upload through their real egress paths. Expect the destination to receive neither sample and the applicable enforced policy to appear in network traffic logs, with a corresponding DLP match in Purview telemetry.
A passed browser test tells you about that browser route. Repeat the check from the agent’s actual path before trusting the same control there.