What Is Data Loss Prevention in the Age of AI?

Data loss prevention is a set of policies and tools that helps detect and prevent data leakage and other forms of risky data sharing. DLP protects sensitive data as people create, store, use, copy, upload, and send it.

DLP has long covered email, endpoints, web traffic, cloud storage, SaaS apps, and collaboration tools. Those channels still matter, but in the age of AI, DLP must also cover prompts, uploads, model inputs, model outputs, and AI agent actions.

For example, a finance employee may try to send a spreadsheet with customer payment data to a personal email account. DLP can detect payment card data and then block, encrypt, quarantine, or alert on the message.

DLP protects sensitive data that could create business, financial, legal, or compliance risk if exposed, including:

• Intellectual property, such as source code, product plans, designs, and trade secrets.
• Personally identifiable information, such as names, addresses, ID numbers, and employee records.
• Protected health information, such as patient notes, diagnoses, claims, and treatment data.
• Payment card and financial data, such as account numbers, payroll files, invoices, and transactions.
• Credentials, secrets, tokens, API keys, certificates, and access data.
• Regulated records, contracts, board materials, and legal documents.

In AI workflows, the same data can appear in new places. It can be pasted into a GenAI prompt. It can be uploaded to a summary tool. It can be indexed by a copilot, pulled into a RAG workflow, or processed by an AI agent.

DLP works by identifying sensitive data, evaluating it against policy, and taking action when risk is detected.

Common DLP actions include:

• Detect: Identify sensitive data in files, messages, prompts, uploads, and repositories.
• Alert: Notify security teams, data owners, or compliance teams about risky activity.
• Block: Stop unauthorized data sharing, upload, copy, paste, or transmission.
• Encrypt: Protect sensitive data before it leaves an approved environment.
• Quarantine: Hold risky files or messages for review before release.
• Redact: Remove or mask sensitive data from a message, file, prompt, or output.
• Coach: Warn users in real time and guide them toward safer behavior.
• Audit: Record policy violations, decisions, and user actions to support investigations, incident response, and compliance review.

DLP works best when it evaluates both content and context. For example, the same file may be appropriate to share with an approved legal team but risky to upload to an unapproved public AI tool.

AI changes data loss prevention because sensitive data no longer moves only through email, endpoints, cloud apps, and file-sharing tools. It now flows through prompts, uploads, model inputs and outputs, RAG workflows, AI agents, and other AI-powered systems. At the same time, AI can surface overshared information, accelerate data access, and expose sensitive content in ways traditional human-driven workflows never could.

Prompts create a new data channel

Prompts are now a data channel. Employees routinely paste source code, customer information, contracts, and other sensitive data into AI tools to speed up their work, creating a new path for data exposure.

DLP should detect sensitive data in prompts, uploads, and copy-and-paste activity while coaching users toward approved AI tools.

AI agents create new access risks

AI agents can act with autonomy by retrieving data, calling tools, querying systems, summarizing files, sending messages, or triggering workflows. If an agent has excessive permissions, weak guardrails, or access to overshared repositories, it can expose sensitive data at scale.

For example, an AI agent with access to internal documentation could retrieve confidential product plans and surface them in a user-facing workflow. This is why access governance matters. Modern DLP should combine least-privilege access, data classification, data lineage, agent monitoring, and behavioral analytics to show not just what data moved, but who or what moved it, why, and where it went.

AI outputs can become exposure points

AI outputs can expose data when they repeat, summarize, infer, or reveal sensitive content. An AI-powered copilot may surface overshared files, a RAG workflow may return private data to the wrong user, and a custom LLM may generate responses based on sensitive training or tuning data. Because AI transforms information before presenting it, the output may appear new even when it originates from sensitive data.

Traditional DLP risk vs. AI-era DLP risk

AI creates new data exposure risks by giving people and agents new ways to access, change, and share information. DLP should help reduce these risks by finding sensitive data, enforcing policy, and giving teams clear visibility.

Proofpoint research shows these risks are already affecting organizations: 42% report a suspicious or confirmed AI-related security incident. Among those organizations, 67% saw threat activity in email, 57% in SaaS or cloud apps, and 53% in AI assistants or agents.

Modern DLP should help reduce:

• Shadow AI and unsanctioned tool use
• Sensitive data in GenAI prompts and uploads
• Overprivileged AI applications and agents
• Sensitive data in training, tuning, embeddings, and RAG workflows
• Prompt injection and adversarial prompting that attempts to extract protected information
• Compromised plug-ins, APIs, OAuth permissions, and connected applications
• Insider risk from careless, compromised, or malicious users

Shadow AI and unsanctioned tools

Shadow AI occurs when employees use AI tools that haven't been approved by the organization. Because these tools may not follow company requirements for data handling, organizations often lack visibility into what data is uploaded, how long it's stored, or how it's used.

For example, a healthcare employee might use an unapproved AI tool to summarize patient notes. DLP can detect sensitive data moving into these tools and guide users toward approved alternatives instead of simply blocking their work.

Overprivileged AI and excessive agency

Overprivileged AI apps and agents can access more data than they need. For example, a copilot may index overshared files, or an OAuth app may receive broad permissions that allow an agent to access multiple systems.

Excessive agency means an agent has too much freedom to act. This raises risk because an agent may not only view data. It may move, summarize, copy, or send it.

DLP should work with access mapping, labels, least privilege, and agent monitoring. Teams need to know what an AI tool can reach and whether that access fits the task.

Insider risk in AI workflows

AI workflows can make insider risk more difficult to pinpoint. Careless users may upload confidential files to public tools, while compromised or malicious users may use AI to locate, summarize, or extract sensitive information at scale.

Context is critical. An engineer summarizing public documentation may pose little risk, while uploading source code to an unapproved AI tool should trigger a different response. Insider threat management helps make that distinction by connecting user behavior with data sensitivity and AI activity.

Modern DLP must protect sensitive data across people, applications, and AI agents. It should integrate with data discovery, classification, DSPM, access governance, insider threat management, data lineage, audit trails, and AI governance.

Visibility across people, apps and agents

Security teams need visibility into how sensitive data moves across users, applications, AI tools, and agents. The goal is to answer questions like:

• Who accessed sensitive data?
• Was the user, application, or agent authorized?
• What sensitive data was involved?
• Was it copied, uploaded, summarized, shared, or included in a prompt?
• Did the activity involve an approved AI tool?
• Did the output expose sensitive information?
• Is this behavior normal for the user or role?

Classification before enforcement

Effective DLP starts with classification: knowing what data is sensitive and where it resides. DSPM helps identify sensitive data across cloud and on-premises repositories, including files with overly broad permissions, so DLP can apply controls before users, applications, or AI agents move that data into risky channels.

For example, Microsoft 365 Copilot can access SharePoint and OneDrive content. If sensitive files are overshared, Copilot may surface them to unauthorized users. DSPM identifies the exposure, while DLP governs how that data is accessed, shared, and used.

Real-time controls and user coaching

Modern DLP should balance security with productivity by applying controls based on context, not blanket restrictions.

Key capabilities include:

• Discover and classify sensitive data before AI tools access it.
• Continuously monitor public GenAI, enterprise copilots, custom LLMs, and AI agents.
• Detect sensitive data in prompts, uploads, files, and outputs.
• Apply adaptive controls such as blocking, encryption, redaction, quarantine, and user coaching.
• Govern access using least-privilege principles.
• Correlate user behavior, data sensitivity, AI activity, and destination risk.
• Maintain audit trails and track data lineage for investigations and compliance.

Real-time coaching helps users make safer decisions at the moment of risk by explaining why an action is unsafe and directing them to an approved workflow.

DLP is one part of an effective AI data governance program. To be effective, it should reinforce clear policies, appropriate access controls, user education, and ongoing oversight.

Define acceptable AI use

Organizations should define what AI use is allowed, what is restricted, and what needs approval. Rules should be clear enough for users to follow and specific enough for security teams to enforce.

For example, a legal team may want GenAI to summarize contracts. A policy may allow summaries of nonconfidential text. It may block uploads that include customer PII unless the tool is approved and governed.

Acceptable use policies should cover:

• Which AI tools are approved
• What data may be used
• Which use cases need review
• Which roles can access each tool
• How to handle prompts, uploads, outputs, and generated content
• How exceptions are approved

Enforce policies with technical controls

Policies need technical controls. DLP helps turn AI governance into action by monitoring data movement and enforcing rules across users, apps, and agents.

Use this process to align DLP with AI governance:

• Discover sensitive data across cloud, SaaS, endpoints, collaboration tools, and on-premises repositories.
• Classify data by risk, business value, and regulatory need.
• Map AI access so teams know which tools, copilots, apps, and agents can reach sensitive data.
• Enforce adaptive policies, such as blocking, redaction, encryption, coaching, escalation, and audit logging.
• Review incidents, user behavior, and audit trails to improve policy over time.

This approach helps organizations support AI adoption while maintaining control of sensitive data.

Drive behavior change

AI governance is not only technical. Employees need to know what data they can use, which tools are approved, and why some actions create risk.

Security awareness training can help teach safe AI use. Real-time coaching can add guidance at the moment of risk, such as when a user tries to paste sensitive data into an unapproved GenAI tool.

Many AI data risks stem from routine work. Users may be summarizing documents, writing code, analyzing data, or trying to work more efficiently. DLP should guide safer behavior without disrupting approved AI use.

Organizations evaluating DLP solutions should look beyond traditional channel coverage. The right approach should protect data across email, endpoints, SaaS, web, cloud stores, public GenAI tools, enterprise copilots, custom LLMs, RAG workflows, and agents.

Use this checklist to assess whether your DLP approach is ready for AI-era data risk:

For example, a financial services firm may want analysts to use approved AI tools but prevent them from uploading customer account data into unapproved services. DLP should support both goals: enabling approved AI use and controlling sensitive data exposure.

AI adoption creates new data paths. Sensitive data can move through prompts, uploads, copilots, custom LLMs, RAG workflows, outputs, and agents. Teams need visibility, labels, access controls, adaptive DLP, audit trails, and user guidance.

Proofpoint offers resources to help teams secure data across people, AI tools, and agents. The Securing and Governing Data for AI guide gives a deeper framework for protecting sensitive data in AI workflows.

Learn how to secure and govern sensitive data across people, AI tools and agents. Read our guide: Securing and Governing Data for AI.