AI assistants and security gaps: a practical approach for securing Microsoft 365 Copilot

This article is part of Oleria's M365 Copilot security series – see more resources here. 

AI copilots have completely upended the conventional tech adoption curve. We’re seeing astoundingly fast uptake across the enterprise world: Microsoft reports nearly 70% of the Fortune 500 now use Microsoft 365 Copilot — and a CNBC survey found roughly 8 in 10 enterprise organizations are using it. Another AI assistant in the Microsoft realm, GitHub Copilot, now totals more than 77,000 enterprise adopters (a 180% year-over-year surge) — accounting for nearly half of all GitHub revenue growth this year.

The rapid adoption of AI copilots is fueled by the promise of tremendous ROI. In fact, a recent IDC study found that companies are already realizing an advantage, with some organizations already reporting 270% return on their AI investments.

But dig a little deeper, and the implementation of AI copilots doesn’t look quite so smooth. CNBC’s survey also showed only half of those enterprises that adopted Microsoft 365 Copilot have deployed the tool to all employees. A Gartner report found even more concerning numbers: Just 6% of companies they studied had moved their Copilot projects into large-scale deployments — and only 3.3% of IT leaders reported that Copilot has generated value for their companies.

Data security challenges slow adoption — and hold back business value

The biggest source of friction? The multitude of data security challenges presented by AI copilots. A Gartner report found 40% of enterprise copilot deployments face delays due to data security concerns¹, while 2 out of 3 CISOs (66%) in a report from Tines said that data privacy is a major barrier to successful AI adoption. 

At the heart of these concerns lies a fundamental paradox: the very capabilities that make AI copilots so powerful also create significant security risks. These tools excel at discovering and leveraging every accessible resource, finding subtle connections between data with unprecedented speed and contextual awareness. However, this strength becomes a critical vulnerability when coupled with the rampant over-provisioning and misconfigured permissions common in most enterprises. 

An AI copilot, given unchecked access to vast amounts of data, can unintentionally surface sensitive information, transforming existing access issues into active security threats. The result? Unintended exposure of highly sensitive, confidential, and valuable data. This exposure can lead to severe consequences, including costly data leaks, stringent compliance violations, and the potential loss of competitive intelligence, ultimately damaging an organization’s reputation and bottom line.

The path forward: The TRUST security approach for securing AI copilots

Despite these risks, organizations cannot afford to put the brakes on adoption of AI copilots. Rather, they recognize the critical competitive advantage to be gained by charging ahead. So, it’s vital to address the inherent identity security challenges through a strong and practical approach that takes a modern angle on protecting sensitive data from over-permissioning and over-exposure risks associated with AI copilots.

T - Thorough identity visibility across all systems

• Complete mapping of user and service account permissions: Get granular around every user and service account's access rights across all platforms. This involves comprehensive scanning and cataloging of permissions, roles, and entitlements, ensuring no shadow access remains hidden.
• Unified view of access rights across siloed systems: Break down identity silos by aggregating access data from disparate systems into a single, centralized view. This provides a holistic understanding of who has access to what, regardless of the underlying infrastructure.
• Discovery of hidden access paths AI could exploit: Analyze access relationships and dependencies to uncover potential pathways that an AI copilot could utilize to access sensitive data. This includes identifying indirect access routes and complex permission chains.
• Identification of orphaned accounts and unused permissions: Proactively detect and eliminate dormant or orphaned accounts and unused permissions. This reduces the attack surface and minimizes the risk of unauthorized access through abandoned credentials.

R - Right-sized access implementation

• Removal of excessive permissions: Implement a systematic approach to identifying and removing unnecessary permissions. This involves conducting regular access reviews and applying the principle of least privilege to minimize the potential impact of a security breach.
• Implementation of least privilege principles: Enforce the principle of least privilege across all systems and applications, granting users and service accounts only the minimum level of access required to perform their tasks.
• Group membership cleanup and refinement: Regularly review and refine group memberships to ensure that users have only the necessary access privileges. This includes removing inactive members and consolidating redundant groups.
• Just-in-time access for sensitive operations: Implement just-in-time (JIT) access controls for sensitive operations, granting temporary permissions only when needed. This minimizes the risk of persistent over-provisioning and reduces the window of opportunity for attackers.

U - Understanding AI permission boundaries

• Configuration of AI-specific service account limitations: Define and enforce strict access limitations for AI service accounts. This includes restricting the types of data that AI tools can access and limiting the actions they can perform.
• Granular control over what data AI tools can access: Implement fine-grained access controls that allow you to specify precisely which data elements AI tools can access. This includes leveraging data classification and tagging to enforce access policies.
• Permission segregation for different AI functions: Segregate permissions based on the specific functions of AI tools. This ensures that each AI function has only the necessary access privileges, minimizing the risk of lateral movement in case of a breach.
• Data classification integration with access controls: Integrate data classification systems with access control mechanisms. This allows you to automatically enforce access policies based on the sensitivity and classification of data.

S - Surveillance of AI access patterns

• Monitoring AI service account behavior: Continuously monitor the behavior of AI service accounts to detect anomalous activity and potential security threats. This includes tracking access patterns, data usage, and API calls.
• Detecting unusual data access requests: Implement anomaly detection algorithms to identify unusual data access requests from AI tools. This includes flagging requests for sensitive data that are outside of normal operating parameters.
• Tracking sensitive data interactions: Maintain a detailed audit trail of all interactions between AI tools and sensitive data. This includes logging access requests, data modifications, and data transfers.
• Identifying potential exfiltration or exposure events: Implement data loss prevention (DLP) mechanisms to detect and prevent potential exfiltration or exposure of sensitive data by AI tools.

T - Targeted remediation of access vulnerabilities

• Automated closure of security gaps: Automate the process of remediating access vulnerabilities. This includes automatically revoking excessive permissions, disabling dormant accounts, and enforcing access policies.
• Guided workflow for fixing over-provisioned permissions: Provide guided workflows that help security teams to quickly and easily fix over-provisioned permissions. This includes providing context and recommendations for remediation actions.
• Contextual risk scoring to prioritize remediation: Implement risk scoring mechanisms that prioritize remediation efforts based on the potential impact of vulnerabilities. This ensures that the most critical risks are addressed first.
• Continuous adaptation as AI capabilities evolve: Adapt and update security measures as AI capabilities evolve. This includes monitoring emerging threats and incorporating new security best practices.

Conventional security tools won’t get you there

Traditional security tools, technologies and strategies are struggling to keep up with the rapidly evolving challenges presented by AI assistants. In short, these conventional solutions fail to provide the comprehensive visibility needed to uncover, remediate, and mitigate the risk of oversharing:

Traditional security approaches are ill-equipped to handle the data exposure risks posed by AI assistants. Data protection tools primarily focus on classification and policy, struggling to keep pace with the speed at which AI can surface and connect sensitive information. Similarly, traditional IAM solutions lack the comprehensive, end-to-end visibility needed to understand and manage access across complex, hybrid environments, often forcing security teams to manually piece together disparate data. Native platform security controls can create silos and blind spots, particularly when data is accessed in non-traditional ways or leaves the platform.

Relying solely on these conventional tools leaves organizations vulnerable to unintended data exposure through over-provisioning and misconfigured permissions. The rapid data discovery capabilities of AI assistants amplify these existing weaknesses, highlighting the need for a more integrated and dynamic security strategy that prioritizes comprehensive visibility, real-time monitoring, and adaptive access controls across the entire digital estate.

How Oleria uniquely addresses M365 Copilot security challenges 

Oleria was born from our firsthand experience as security operators, witnessing the critical visibility and control gaps that plagued traditional identity security. Even as GenAI began its rapid ascent, we recognized the inherent danger of rampant over-provisioning and outdated tools, which were making identity the primary source of enterprise risk.

Today, Oleria's Trustfusion Platform and Oleria Identity Security are uniquely positioned to address the challenges of securing AI copilots, providing CISOs and security teams with the necessary visibility and control to enable agility while safeguarding data security and privacy — giving you unique capabilities:

• Unified identity visibility

Oleria consolidates access permissions from all IAM and SaaS applications, providing a single, comprehensive view down to the individual resource level. By automatically surfacing unused accounts and identifying excessive privileges, Oleria eliminates identity silos and provides a holistic understanding of potential vulnerabilities across hybrid environments.

• Automated least privilege enforcement

Oleria simplifies least privilege enforcement by automatically identifying and remediating over-provisioned access. It facilitates the removal of unused accounts and enables a shift from broad, role-based permissions to targeted, individual-level access, ensuring only necessary access is granted, including mitigating over access for AI tools.

• AI-ready architecture

Oleria's architecture is designed to secure AI assistants with granular control over AI service account limitations and precise configuration of data access permissions. Capabilities for monitoring AI access patterns and detecting anomalies ensure AI tools operate within defined boundaries, preventing unauthorized data access.

• Integration with Microsoft 365

Oleria extends comprehensive visibility and control to Microsoft 365 Copilot environments, enabling organizations to confidently deploy this platform while mitigating associated data security risks. This integration proactively prevents unintended data exposure within the Microsoft 365 ecosystem.

Turning AI security into your competitive advantage

Every enterprise today feels incredible top-down pressure — from investors, the board, and C-suite leadership — to accelerate toward a Copilot-enabled future. Regardless of risks and challenges, it’s clear that organizations that can deploy AI copilots effectively and securely stand to gain tremendous competitive advantage in their markets. 

But the hype and urgency is leading to a familiar bias toward speed over security: deploy quickly and figure out security later. The reality is that the most valuable applications of AI copilots will require businesses to fully trust their AI assistants. That makes security the foundation, rather than the post hoc patch.

By following the TRUST approach for securing AI copilots, organizations can lay that foundation — and avoid ending up among that 40% of AI projects delayed by data security. This gives an enterprise the agile head-start to move fast and enable powerful value-creation applications — while confidently protecting all the valuable information within the enterprise.

Learn more about identity security in the AI age: Read our in-depth analysis on Microsoft 365 Copilot access control risks