Our recommendation is to first identify and address any underlying security concerns that could result in harmful consequences when deploying AI into your environment.

1. AI is only as secure as the data it can access

AI operates within the context of the data it has access to, utilising the permissions of the person using it. As a result, any information that is incorrectly secured, overshared, or misclassified could be inadvertently incorporated into AI’s outputs. The consequences of this may lead to the disclosure of confidential information that results in the failure to comply with regulatory requirements.

2. AI can amplify existing security gaps for attackers

AI can enhance workforce efficiency. However, it can also present opportunities for attackers to operate more effectively if they gain access to your environment. Malicious actors may utilise AI to accelerate the exfiltration of sensitive information, which could then be used for extortion or to escalate their attack.

3. AI requires relevant and high-quality data sources to provide maximum effectiveness

Although not strictly security, data governance should also be considered as part of an AI strategy. The effectiveness of AI is determined by the quality and relevance of the data it draws upon. Feeding AI systems with clean, current, and contextually appropriate information ensures that generated outputs are accurate and trustworthy. Redundant, obsolete, or trivial data, often referred to as ROT, on the other hand, can distort results, propagate errors, and undermine confidence in AI-driven insights.

Note: While these steps apply to any AI model (Gemini, Microsoft 365 Copilot, ChatGPT, etc.), the tools and functionalities discussed below are specific to Microsoft 365 environments.

A crucial step in safeguarding your organisation’s information is the regular review and adjustment of permissions. Broad or unnecessary access increases the risk that sensitive or confidential content may be inadvertently exposed to AI systems. Periodic audits and revocations of access rights help maintain relevant permissions that follow the principle of least privilege, helping to significantly reduce the likelihood of oversharing.

Intelogy offers a Copilot Readiness Assessment for Microsoft 365 that can also help with this step.

Microsoft provide various tools to help you in implementing this:

SharePoint Advanced Management

One of the included components is data access governance reports that enable you to understand potential oversharing by providing a list of OneDrive and SharePoint sites that have had the most new sharing links created in the last 30 days, as well as information about sites with the greatest number of items with access to everyone except externals.  

Another component is the site access reviews. They allow administrators to delegate the process of reviewing data access governance reports to site owners of overshared sites. This can be beneficial as administrators don't always have the knowledge or access to determine and address sharing issues for a site compared to the site owners. Site owners will receive an email with comments from an administrator and can then use the links provided to get an overview of the reported access and the ability to modify the access. 

Microsoft Entra ID Governance

Entitlement management can be used to automate the request, approval, and lifecycle process for access to groups and SharePoint sites. This capability empowers users to request access to the assets they need, whilst allowing approvals to be used as a prerequisite before granting access. You can also define durations for when access should expire or require renewal.

Access reviews provide the ability to regularly review and manage group memberships to make sure that the right people have access. This can be delegated to group members, group managers, group owners, or select individuals.

Classifying* your data is important not only to make it easier to understand the sensitivity of information but also to enforce persistent protection** that can secure the data even if it leaves your controlled environment.

*Classification is the act of labelling content in conjunction with a data classification scheme for your organisation.

** Persistent protection is the act of enabling additional controls that persist with the content no matter its location or integration.

The following tools are available as part of Microsoft Purview to support this requirement:

Sensitivity labels

Sensitivity labels primarily offer a method to classify content, including emails and files. For files, sensitivity labels can apply not only to those stored in Microsoft 365 but also to files located in file shares or other cloud apps. Classification is important because it helps identify the sensitivity of specific content and can serve as a condition for tools such as data loss prevention.

Beyond classification, sensitivity labels can also be used to apply encryption, which allows you to apply granular access control that remains persistent with the content, no matter the environment it ends up in. Persistent access control can help add a layer of protection when file location/repository permissions are overexposed. Encryption enforced by sensitivity labels relies on Azure Rights Management (Azure RMS), which is embedded in Microsoft services. Third-party apps, including AI solutions, are unable to read content that is protected by Azure RMS.

Data loss prevention (DLP)

Data loss prevention can be used to protect content that has been classified using sensitivity labels. Regarding AI, one example is the ability to use Endpoint DLP to restrict the ability to upload files using specific sensitivity labels to AI sites you define. This is ideal when you have sensitivity labels that are not applying encryption, which may be required when open collaboration is a priority.

Microsoft has also enabled the ability to create DLP policies that prevent Microsoft 365 Copilot from interacting with content using specific sensitivity labels. This feature is currently in public preview but is soon to be rolled out for general availability. 

Making sure that access to your environment is and remains secure is key for many different reasons, but in the age of AI, this is all the more important. Adopting a defence-in-depth approach will strengthen your overall protection.

Two important practices to implement are strong authentication and conditional access:

• Strong authentication involves requiring people accessing your environment to use phishing-resistant authentication. Phishing-resistant authentication is an advanced version of MFA that reduces the impact of phishing attacks and provides far greater security than traditional MFA. To learn more about why standard MFA is no longer enough, please see our previous blog post.
• Conditional access is used to evaluate various signals involved in the authentication process, and depending on the met criteria, access can be granted or denied. This is effective when you need to evaluate whether a device is managed or compliant to help combat malicious attackers utilising unknown devices.

Connectivity between apps is heavily utilised across many organisations, but it is important to make sure that this is enabled with proper governance and only for the applications that you approve.

If you’re using Entra ID or Microsoft 365 as your primary platform for identity management, you should review the way that app consent is configured to avoid users being able to grant access to unapproved apps.

For example, ChatGPT has the option to connect to your organisation’s Microsoft 365 environment and access files stored in your users’ OneDrive and SharePoint sites.

If you do not have app consent permissions correctly configured, this can result in users being able to leverage features like this, which effectively provides ChatGPT access to use your data in its analysis and responses.

By proactively monitoring AI usage across your organisation, you gain valuable insights into where, how, and what data is being accessed or processed by artificial intelligence. This enables you to identify emerging patterns, potential risks, and points of non-compliance. Establishing a comprehensive view of AI activity helps ensure responsible adoption, supports compliance requirements, and reduces the likelihood of data misuse while strengthening your organisation’s overall security posture.

Data Security Posture Management (DSPM) for AI is a useful tool available as part of Microsoft Purview that enables you to monitor the usage of AI across your organisation and understand what types of sensitive information are being utilised. Although this is a Microsoft tool, this includes the usage of third-party AI tools such as ChatGPT, Gemini, Claude, and many more.

Data lifecycle management is pivotal in enhancing the reliability of AI. By making sure that data is subject to automated retention and disposal, you minimise the risk of outdated or redundant data influencing AI’s recommendations.

Microsoft Purview offers advanced retention capabilities, allowing organisations to automatically retain data for compliance purposes while managing secure and timely disposal when required.

Security isn’t just technical, it’s cultural. Building robust defences goes beyond configuring systems and enforcing policies; it demands an investment in your people.

Training users to recognise potential risks and empowering them to act as stewards of organisational information is just as important as deploying the most advanced security tools. This means equipping team members with the knowledge to apply appropriate sensitivity labels, set the correct permissions, and understand the consequences of improper sharing. It also involves ongoing education about the latest threats, common pitfalls, and best practices for data handling in an AI-powered environment.

Cultivating a culture of security awareness transforms your workforce into a proactive line of defence that complements technological safeguards.

Traditional security measures are simply not enough anymore to handle the unique complexities of AI. Organisations must move beyond basic compliance. If you are ready to deploy AI into your workplace, we are here to advise you on how to make your security AI-ready.

In 2025, Intelogy officially became a Ready Tier Partner in Microsoft’s highly exclusive Copilot Jumpstart Partner Initiative. We are well-equipped to help any organisation, big or small, start their AI transformation journey in a secure and governed manner. This truly sets us apart.

Microsoft offers a free assessment to evaluate your current AI security across four key pillars: Prepare, Discover, Protect, and Govern.

After you complete all the questions, you will instantly receive your results, which indicate the status of your security posture: Initializing, Foundational, or Optimized. You can review your results and download the full report for details and recommendations to strengthen your organisation’s security posture for Generative AI.