AzureFeeds

Why AI Changed the Data Security Problem

AI does not create entirely new categories of risk—it supercharges existing ones. Traditional data leakage stems from ordinary behavior: sharing a document too broadly, sending an email to the wrong person, copying regulated data to an uncontrolled device. Generative AI amplifies all of these because of the power and speed with which it can proactively surface content that may be obsolete, over-permissioned, or ungoverned. DSPM exists to help with exactly this challenge: it continuously scans your environment to identify sensitive data, assess risk, and recommend actions to reduce exposure.

• Oversharing at Scale
  Before AI, an overshared SharePoint file might sit unnoticed. Now, Copilot can summarize it in response to a casual prompt, distributing its contents far beyond the original audience.
  
  


• Prompt Leakage
  Users can inadvertently expose sensitive information—financial account numbers, health records, project code names—simply by typing them into a Copilot prompt. Because AI interactions feel conversational, users tend to drop their guard.
  
  


• Shadow AI
  Beyond sanctioned tools, employees experiment with unapproved AI services.
  
  


• Autonomous Agents
  
  

  Autonomous agents expand the data security threat surface by acting independently on sensitive information across systems and boundaries. Their ability to access and share data without direct user interaction increases the risk of oversharing, exfiltration, and unauthorized access, while also introducing complex behavior patterns that are harder to monitor, govern, and control using traditional security models.

  
  

What Microsoft Purview Now Brings Together

Data Security Posture Management (DSPM)

DSPM consolidates insights from Data Loss Prevention (DLP), Insider Risk Management, Information Protection, and Data Security Investigations into a single view for monitoring data risks, policy coverage, and posture trends. Now also in Public Preview, DSPM extends coverage to third-party SaaS and IaaS platforms such as Google Cloud Platform, Snowflake, and Databricks, and integrates with partner solutions including Cyera, BigID, and OneTrust for comprehensive risk insights.

A central innovation in this version is data security objectives—prominent, selectable cards that each represent a specific security goal. Selecting an objective guides administrators through an end-to-end workflow that groups together the most relevant Purview solutions—information protection, DLP, Insider Risk Management, and eDiscovery—so teams can focus on achieving a specific data security outcome rather than navigating separate solutions.

Each Outcome card displays key metrics such as the percentage of data covered by policies, the number of risky sharing incidents, and improvements over time. Within each outcome, DSPM surfaces suggested prioritized actions—applying sensitivity labels, configuring DLP policies, or investigating alerts—all tailored to the organization’s data. Administrators can take action directly from the workflow, including remediating oversharing, configuring one-click policies, or launching investigations into suspicious activity.

DLP Integration for AI Interactions

DLP is one of the core solutions integrated into DSPM’s unified approach. The Activity Explorer’s AI activities tab captures events where DLP rules were matched during AI interactions—including prompts, responses, and browsing to generative AI sites. DSPM can automate remediation steps such as removing public sharing links or applying data loss prevention policies to help prevent incidents before they happen.

AI Observability and Agent Governance

Dedicated dashboards and metrics monitor risks associated with AI apps and agents. AI observability enables tracking of agent-specific activities—oversharing, exfiltration, and unusual access patterns—across both Microsoft and third-party environments. Enhanced reporting provides advanced filtering and customizable views, supporting granular analysis of sensitive data usage, DLP activity, and posture trends. Audit logs and activity explorer features help track interactions with AI apps and agents, supporting compliance investigations and incident response.

AI-Powered Security Operations

DSPM not only secures and governs AI apps and agents but also uses Microsoft Security Copilot and AI agents to help secure and govern data. AI analyzes access patterns, sharing behaviors, and policy gaps to surface actionable risks and can detect unusual activity such as excessive sharing or suspicious downloads. Under administrator guidance, AI agents can take direct action on detected risks—removing public sharing links, applying DLP policies, or revoking permissions. These actions are always audited. To streamline investigations, AI-driven triage agents review alerts from DLP and Insider Risk Management solutions, filtering out noise and highlighting the most critical threats.

Three Practical Starting Points

For many organizations adopting generative AI, the biggest hurdle isn’t recognizing new risks—it’s figuring out where to begin. A “boil the ocean” approach can stall progress, while tackling a few targeted areas delivers quicker wins.

The best early moves are those that reduce exposure quickly, improve visibility, and build a foundation for stronger governance over time.

Starting Point 1: Enable prompt-level protection for Microsoft 365 Copilot

An effective first step is to put guardrails on the prompts users enter into AI. Microsoft Purview DLP allows administrators to restrict Microsoft 365 Copilot and Copilot Chat from processing prompts that contain sensitive information. In practice, users are often more comfortable pasting data into a chat prompt than attaching it to an email, which means a well-meaning employee could inadvertently feed a confidential file or personal data into Copilot.

Enabling prompt-level DLP creates an immediate safety net: if a user’s prompt includes, say, a credit card number or a customer’s national ID, Copilot will detect it and refuse to process or share that content. DSPM provides suggested prioritized actions—including configuring DLP policies—that can be activated directly from the workflow, and recommended policies can start in simulation mode. Simulation mode lets you see what would have been blocked or flagged, without actually interrupting users, so you can fine-tune the policy and prepare your helpdesk for any questions. Once you’re comfortable with the results, switching to enforcement mode will actively block disallowed prompts and log those events for review. 

By activating this one control, you’ve significantly reduced the most immediate oversharing risk—the “oops, I pasted the wrong data” scenario—within hours of starting your AI governance program.

Tradeoff: Simulation mode provides safety but delays enforcement. For organizations with imminent regulatory exposure, consider shortening the simulation window and monitoring alert volumes closely.

Starting Point 2: Gain visibility into shadow AI usage before broad enforcement

The second step is to illuminate what’s happening in the shadows. Before rushing into blocking every unsanctioned AI tool, it’s crucial to understand how and where AI is being used across the organization. In most enterprises, there’s an official layer of AI usage and an often larger, unofficial layer—employees experimenting with free online AI chatbots, writing assistants, or code generators.

DSPM provides this visibility. The Discover > Apps and agents dashboard shows AI apps used across the organization, including the top 20 most recently used agents, with details about sensitive data they accessed and how they are protected by Purview policies.

The AI observability page provides a broader inventory of all AI apps and agents with activity in the last 30 days, including how many are high risk and the total with sensitive interactions. The Activity Explorer’s AI activities tab shows when users browsed to generative AI sites, the prompts and responses involved, whether sensitive information was present, and whether DLP rules were matched. Armed with this insight, you can make informed decisions. If you discover that the majority of “AI consumption” comes from just two external apps, you might focus your immediate controls on those two. Conversely, if the data shows most unsanctioned usage is low-risk, you might decide to monitor rather than block it.

The key is visibility first, enforcement second—letting real data guide where to tighten controls versus where to offer secure alternatives.

Tradeoff: Visibility without timely follow-through can create a false sense of security. Set a defined window (e.g., 30 days) after which findings must translate into at least one concrete policy action.

Starting Point 3: Operationalize DSPM objectives for Copilot

A stronger third starting point is to use DSPM as your operational guide, not just a dashboard of charts. DPSM introduces data security objectives—each one a focused end-to-end workflow for a specific outcome. Rather than configuring individual features in isolation, you select an objective and let Purview navigate you through achieving that outcome with the relevant tools.

For generative AI, the key objective to leverage early is “Prevent data exposure in Microsoft 365 Copilot and Microsoft Copilot interactions”. By selecting this objective in the Purview portal, you’re effectively telling Purview, “help me implement whatever is needed to make Copilot safe with our data.” The DSPM interface then groups together the critical pieces: it may prompt you to enable a DLP policy, suggest applying or refining sensitivity labels on content, or surface an Insider Risk Management policy template for detecting AI-related risky behavior. It also surfaces metrics so you can track progress—for example, the percentage of data covered by policies, or the number of risky sharing incidents that have been remediated.

Using DSPM objectives keeps your team aligned on a clear goal from day one. It shifts the conversation from “what knobs do we turn on?” to “how do we achieve this outcome?” You follow a guided plan curated by the platform’s intelligence rather than navigating five different admin pages and hoping it adds up to protection.

Tradeoff: Objectives streamline the path but can obscure the underlying complexity. Teams should periodically step outside the guided workflow to review the full policy landscape and ensure no coverage gaps exist between objectives.

From Visibility to Remediation: Turning Insights into Action

Automated Remediation at Scale

DSPM can automate remediation steps such as removing public sharing links or applying data loss prevention policies to prevent incidents before they happen. Under administrator guidance, AI agents within DSPM can take direct action on detected risks—removing sharing links, applying DLP policies, or revoking permissions—and these actions are always audited. This moves the operating model from manual, one-at-a-time fixes to systematic, policy-driven remediation.

Closing the Loop: From Risk to Standing Policy

DSPM’s data security objectives surface suggested prioritized actions such as applying sensitivity labels, configuring DLP policies, or investigating alerts, all tailored to the organization’s data. Reporting and analytics are organized by outcome, making it easier to identify and report improvements, compliance, and risk reduction. This turns recurring findings into standing preventive controls. Instead of re-running assessments and manually fixing the same patterns, administrators create durable policies that enforce the desired state going forward.

Alert-Driven Investigation and Tuning

Audit logs and activity explorer features help track interactions with AI apps and agents, supporting compliance investigations and incident response. Integrated investigation and forensics tools support rapid incident response and root cause analysis for data security events. Impact prediction visuals and progress tracking for remediation steps are surfaced throughout DSPM, enabling administrators to quantify the effect of their actions and adjust course.

The closed-loop process is: Discover (DSPM scans and risk assessments) → Remediate (automated actions and bulk fixes) → Prevent (create or tighten DLP and auto-labeling policies) → Monitor (alert review, investigation, and policy tuning).

What “Good” Looks Like in a Regulated or Risk-Aware Organization

A mature AI governance posture is defined by measurable outcomes and sustainable operating rhythms—not feature count:

• Clear, communicated AI usage policies. Users know what is and is not acceptable in AI interactions because the tools reinforce the rules. DLP policy tips delivered at the moment of a violation are a primary training mechanism—they remind users in context why their prompt was blocked and what to do instead.
  
  


• Measured enablement over blanket bans. Leading organizations allow Copilot with appropriate controls and restrict only truly unacceptable scenarios. Policies deployed initially in simulation mode provide data to calibrate enforcement thresholds before blocking. This avoids productivity backlash while preserving security posture.

• High data hygiene and classification rates. Purview’s AI protections depend heavily on sensitivity labels. If everything is unlabeled or “General,” label-based controls have nothing to act on. Mature organizations invest in auto-labeling and mandatory labeling to close this gap before deploying AI at scale. DSPM’s data security objectives include suggested actions such as applying sensitivity labels, directly tying classification to governance outcomes.

• Quantifiable risk reduction. Security leadership can produce metrics from Purview that show trend lines: DSPM Outcome cards display the percentage of data covered by policies, the number of risky sharing incidents, and improvements over time. These figures feed directly into compliance reporting and audit evidence. Key metrics are tracked over time, supporting continuous improvement of the organization’s data security posture.

• Cross-functional governance. AI governance is not a solo IT Security effort. Stakeholders from security, compliance, legal, and business units review AI usage patterns, discuss policy tuning, and evaluate new Purview capabilities as they release. Role-based access controls within DSPM provide granular access to features and AI content for delegated administration and compliance, enabling this cross-functional model without overexposing sensitive data to every participant.

Tradeoff: Strict enforcement can frustrate power users and slow AI adoption. Organizations should explicitly define escalation paths—if a legitimate use case is blocked by DLP, there must be a fast process to review and adjust, rather than a permanent “no.”

A Phased Adoption Model

*Phase 1 should take weeks, not months—the objective is to establish a baseline before risk accumulates.

*Phase 2 is where enforcement generates measurable risk reduction.

*Phase 3 is ongoing: as Microsoft continues extending Purview to additional AI apps and agent types, the governance framework must evolve in tandem. 

The DSPM preview’s integration with third-party SaaS and IaaS platforms (Google Cloud Platform, Snowflake, Databricks) and partner solutions (Cyera, BigID, OneTrust) means the governance perimeter can expand alongside the organization’s AI footprint.

Conclusion

AI adoption and data protection are not opposing forces. Microsoft Purview now provides the visibility, policy controls, and remediation workflows to move from discovering AI risk to actively governing Copilot, third-party AI apps, and agents at scale. DSPM surfaces oversharing and AI usage patterns through unified dashboards, data risk assessments, and AI observability. DLP blocks sensitive data in prompts and restricts AI access to labeled content. Insider Risk Management detects adversarial AI behavior. AI-driven triage and remediation agents close the gap between identifying a problem and fixing it—with every automated action audited.

The path forward starts with practical actions: enable prompt-level DLP, illuminate shadow AI usage, and operationalize DSPM’s “Prevent data exposure in Microsoft 365 Copilot and Microsoft Copilot interactions” objective. From there, enforce what you find, measure the results using DSPM’s outcome-based metrics, and progressively mature your governance posture.

Organizations that operationalize this loop will be in a strong position: able to say, “We use AI to work smarter—and we have the safeguards in place to do it safely.”