For organisations using Microsoft 365, Microsoft Purview consulting can help extend existing information protection and data governance controls into these AI scenarios.

What is AI data security?

AI data security is the combination of policies, processes and technical controls used to protect organisational data as it is accessed, processed or shared through AI applications and agents.

It sits within the broader disciplines of AI security and AI governance, but focuses specifically on the data involved: what information AI systems can access, how sensitive information is protected and how people and AI systems interact with organisational data.

AI can also amplify existing security weaknesses. Information that was already overshared or poorly classified within Microsoft 365 can become easier for users and AI systems to discover and use.

Common AI data leakage risks

AI data leakage does not necessarily require a sophisticated attack. In many cases, it starts with normal user behaviour.

Information pasted into public AI tools

An employee pastes a customer list, contract, financial information or other sensitive material into a public chatbot. Once information leaves the controlled environment, the organisation may have limited visibility or control over what happens to it.

Overshared Microsoft 365 content

If an employee already has access to a SharePoint file they should not realistically be able to see, Microsoft 365 Copilot may make that information easier to find and use.

The underlying problem is often not AI itself, but the permissions and data governance that existed before AI was introduced.

Reviewing permissions and data governance should therefore form part of broader Microsoft 365 consulting and AI readiness work.

AI plug-ins and free tools

AI capabilities are appearing inside applications employees already use. Confidential information may be uploaded to free AI services, browser tools or AI-enabled plug-ins without users considering where that information is being processed.

Personal AI accounts

Corporate controls become less effective when employees use personal accounts or unapproved services. AI data protection therefore cannot focus solely on securing the organisation’s approved AI platform.

AI agents

Agents can be given access to organisational data and systems and perform actions on a user’s or organisation’s behalf.

Organisations need to understand what data an agent can access, what actions it can take and who is responsible for it.

Why is shadow AI difficult to control?

Shadow AI is the use of AI applications or services outside an organisation’s approved or managed technology environment.

The first challenge is visibility. If IT does not know which AI applications are being used, it cannot properly assess what information employees are sharing with them or whether those tools meet security requirements.

Blocking every AI service is rarely a practical long-term strategy. A better approach is to discover what is already being used, assess the risk and determine which applications should be sanctioned or unsanctioned.

Microsoft Defender for Cloud Apps can help provide that visibility by discovering cloud applications in use and assessing their risk.

How Microsoft Purview protects data used by AI

Microsoft Purview provides data security and compliance capabilities that can be applied to AI use. Organisations can combine classification, DLP, monitoring and risk management to address different ways information may be exposed.

Sensitivity labels help protect sensitive information

Sensitivity labels allow organisations to classify information according to its sensitivity and apply appropriate protection requirements.

For example, information might be classified as public, internal, confidential or restricted, with different access and sharing rules attached to each.

This becomes particularly important with AI. If sensitive information is poorly classified or widely accessible before AI is introduced, AI can expose those weaknesses much faster.

Microsoft Purview DLP can prevent sensitive data leaving the organisation

Classification tells you what the data is. Microsoft Purview Data Loss Prevention (DLP) helps control what can happen to it.

Depending on the scenario and configuration, DLP policies can:

1. monitor activity involving sensitive information
2. warn users before information is shared
3. block sensitive data from being sent to unsanctioned applications

For AI, this can provide an enforcement layer when employees attempt to send sensitive information to external AI services.

Data Security Posture Management for AI provides visibility into data risk

Data Security Posture Management (DSPM) for AI helps organisations understand AI activity and associated data risks, including potential data leakage, oversharing and inappropriate AI use.

This helps security teams identify patterns of risky behaviour, understand where sensitive information may be exposed and prioritise remediation.

Insider Risk Management helps identify risky AI behaviour

Not every data security incident is malicious. An employee may upload information to an AI tool because it saves time without understanding the security implications.

Microsoft Purview Insider Risk Management can help identify potentially risky user activity and provide security teams with information to investigate it.

Defender for Cloud Apps helps identify and control shadow AI

Microsoft Defender for Cloud Apps addresses another part of the problem: discovering the cloud applications employees are actually using.

For generative AI, IT teams can use this visibility to:

1. discover AI applications in use
2. understand usage and assess risk
3. determine which applications should be sanctioned or unsanctioned

Together, these technologies create layers of protection around organisational data rather than relying on a single AI security control.

What does good AI governance look like?

Technology controls are only part of effective AI governance. Organisations also need clear rules around which AI systems can be used, what information they can access, who approves them and who remains accountable for their operation.

At A1 Technologies, we frame practical AI governance around four areas:

1. Policies: Clear boundaries for AI use, including acceptable use, approved tools, human oversight and data handling.
2. Procedures: Processes for approving AI use cases, conducting risk reviews, managing AI agents and responding to incidents.
3. Management tools: Technical controls such as sensitivity labels, DLP, Insider Risk Management and DSPM for AI that help enforce those policies.
4. Governance and oversight: Regular review of AI risks, incidents, new use cases and outstanding actions.

This needs to be ongoing. New applications appear; employees find new uses for existing tools and agents gain access to additional systems and information.

5 steps to improve AI data security

Organisations do not need to solve every aspect of AI governance at once. A practical starting point is to understand what is already happening and establish some basic controls.

Discover the AI already in use

Identify approved and unapproved AI tools rather than assuming your official Copilot deployment represents your entire AI footprint.

Establish AI registers

Maintain an Approved AI Tools Register and AI Systems Register covering what has been approved, what it is used for, what information it can access and who owns it.

Put an Acceptable Use of AI Policy in place

Give employees practical guidance about which tools they can use and what organisational information can be shared with them.

Review existing data security controls

Assess permissions, data classification, sensitivity labels and DLP. AI can magnify existing data governance problems, so the underlying Microsoft 365 environment needs to be part of AI readiness.

Establish ongoing AI governance

Regularly review AI use, risks, incidents, controls and new applications rather than treating AI governance as a one-off project.

A useful starting point is being able to answer three questions:

1. What AI is being used across your organisation?
2. What organisational data can it access?
3. Who owns each AI tool or system?

If you cannot answer those questions yet, that gives you somewhere practical to begin.

Secure your Microsoft data for AI

As organisations adopt Copilot, third-party AI tools and agents, protecting organisational data requires visibility into how AI is being used, clear governance around approved use and technical controls that protect sensitive information.

A1 Technologies helps organisations assess and implement Microsoft Purview, sensitivity labels, Data Loss Prevention, Data Security Posture Management and related Microsoft security controls as part of a practical approach to securing data for AI.

If you’re assessing how prepared your Microsoft environment is for AI, talk to our Microsoft security and compliance team about your current data security and governance controls.