Forcepoint today launched an AI Data Security platform that extends traditional data loss prevention (DLP) to every prompt, file upload, and response generated by AI services, including Microsoft Copilot and emerging autonomous agents. The platform, available both on-premises and in the cloud, aims to stop sensitive corporate data from leaking through sanctioned and unsanctioned AI tools—a pressing problem for the millions of workers now using generative AI inside everyday apps like Word, Teams, and Outlook.

The Core of the Announcement: DLP Inline, Agent Controls, and Tenant Awareness

The platform doesn’t simply watch AI activity from the sidelines. It inspects content in real time, blocking or restricting sensitive data before it reaches an external service. Forcepoint calls this approach “data-first,” meaning that protection travels with the information itself, regardless of whether it moves through a managed endpoint, a browser, a cloud app, or an AI agent.

Key new capabilities include:

  • Inline prompt and response inspection: Every interaction with Copilot, ChatGPT Enterprise, Claude Enterprise, AWS Bedrock, and other connected services is screened. The system looks for personally identifiable information (PII), payment card data, protected health information, source code, financial records, and trade secrets. If a user types “summarize our Q2 earnings report” into an unapproved tool, the platform can block the prompt, warn the user, or redact the sensitive portion.
  • AI Agent Gateway: Autonomous agents that access business applications (Salesforce, Microsoft 365, Jira) are now governed through least-privileged, field-level controls. Instead of giving an agent broad, persistent credentials, admins can define exactly which fields it can read—for example, letting a sales agent see account names but not payment details. The gateway prevents agents from holding direct application credentials, reducing the blast radius if a prompt injection occurs.
  • Shadow AI discovery and control: The platform identifies unsanctioned AI tools—from browser-based chatbots to unapproved mobile apps—and enforces policy in real time. Admins can allow, block, or conditionally permit services; for instance, letting employees use a public AI tool for general queries but blocking any file upload or paste that contains corporate data.
  • Personal vs. corporate tenant detection: Many employees use both personal and work accounts on the same AI service. Forcepoint can distinguish between the two, allowing access only to the organization’s governed corporate tenant while blocking personal accounts on managed devices. This closes a common gap where a corporate Copilot subscription coexists with a personal ChatGPT login on the same machine.
  • ARIA policy assistant: An embedded AI assistant suggests and refines security policies in plain English. A security lead can ask, “Show me where customer credit card data might leak through AI” and ARIA will propose controls. The idea is to lower the expertise barrier, though all recommendations still require human review before enforcement.

All of this is managed from a single console called “Forcepoint Insights,” which also provides board-ready reports on risk trends, blocked threats, and risky user behavior.

What This Means for You

For Everyday Office Workers

If your company deploys this platform, you may not notice a difference—until you try to share something sensitive with an AI tool. For example, pasting a customer contract into Copilot for a summary might trigger a pop-up warning that the content is restricted. In high-risk cases, the prompt may be blocked entirely. The system does not record your keystrokes; it inspects the data as it moves toward the AI service. You’ll likely be informed with a clear explanation, such as “This file contains PCI data and cannot be uploaded to an unapproved AI tool.”

For personal AI use on a work device, the tenant detection feature means you might be forced to use the corporate version of a service—even if you’re accustomed to your own account. This doesn’t mean the company monitors your personal chats; it simply prevents corporate data from being mixed with an ungoverned consumer account.

For IT Administrators and Security Teams

You gain a single pane to monitor and enforce AI data policies across your entire environment. Instead of juggling separate tools for shadow AI, endpoint DLP, cloud app security, and agent governance, you can now create one set of rules that follow data no matter where it goes. The platform integrates with Microsoft Information Protection (MIP) labels, so existing classifications automatically apply to AI workflows. For instance, a document labeled “Confidential – Finance” in SharePoint will inherit the same restrictions when a user tries to feed it into Copilot.

Deploying the AI Agent Gateway is particularly valuable for organizations automating workflows. You can grant a support agent access to ticket data in Jira but restrict it from viewing customer billing fields. This granularity reduces the risk of exposing data through misconfigured automations—a growing concern as more business processes become agent-driven.

For Developers Building AI Agents

The platform introduces a new way to handle authentication for agents. Instead of baking static API keys into your code, you route agent requests through the gateway, which enforces field-level access. This means you can build an agent that queries Microsoft Graph for a user’s calendar events but is blocked from reading mail content—all without hard-coding sensitive scopes. The result is a more secure, auditable architecture that doesn’t rely on the agent itself being perfectly hardened.

How We Got Here: The Race to Secure Generative AI

It wasn’t long ago that enterprise AI security meant simply blocking all unapproved tools. But that approach crumbled under employee demand. People began pasting sensitive data into free chatbots without IT’s knowledge—a phenomenon Forcepoint and other vendors call “shadow AI.” Early responses relied on visibility dashboards that tracked which AI services were accessed, but that only helped after a leak had occurred.

Microsoft addressed part of the problem by embedding Copilot deeply into Microsoft 365, giving enterprises some control through Purview and sensitivity labels. But those controls often stop at the document level; they don’t inspect prompts or agent actions in real time. Forcepoint’s announcement marks a shift from “knowing” to “stopping.” By extending traditional DLP—already mature for email, endpoints, and the web—to AI interactions, it promises to plug a gap that has widened with every new AI feature in Word, Teams, and Outlook.

First Steps: How to Deploy Without Breaking Productivity

If you’re considering this platform, don’t start by turning on every restriction. A phased rollout preserves trust and keeps business moving:

  1. Inventory your AI exposure. Use the platform’s discovery tools to see which sanctioned and shadow AI services employees actually use. Identify personal vs. corporate tenant use and top data flows.
  2. Classify your sensitive data. Leverage existing MIP labels and DLP classifications. Ensure financial records, customer PII, source code, and regulated documents are properly tagged.
  3. Begin with alert-only mode. Deploy rules in monitoring mode to understand false positives and user behavior. For example, alert on any prompt containing a credit card number, but don’t block it yet.
  4. Enforce for the highest-risk data first. Block clear violations: sending unencrypted credentials to an AI chatbot, uploading an employee performance review to a consumer AI service.
  5. Separate agent governance from user policies. Treat autonomous agents differently. Use the Agent Gateway to enforce least-privileged access and audit every action, even if the agent is deemed “trusted.”
  6. Educate and communicate. Tell users what is allowed, what is blocked, and why. Provide approved alternatives. The goal isn’t to ban AI but to channel it safely.

What to Watch Next

Forcepoint has made ambitious coverage claims, but real-world accuracy will determine the platform’s success. False positives in inline inspection could frustrate users and drive more shadow AI; false negatives could leave sensitive data exposed. Early adopters should test the platform against their actual workflows—especially for nuanced data like unstructured financial text or non-English prompts.

Microsoft’s own security stack (Purview, Defender for Cloud Apps, Copilot governance) will also evolve, potentially overlapping with Forcepoint’s features. For now, the platform offers a complementary layer for organizations that need fine-grained, real-time control—particularly around agent access and shadow AI. If the execution matches the promise, it could become a critical piece of the AI security puzzle for Windows and Microsoft 365 shops.