Bing DLP Update Blocks Copilot Access to Sensitive Office Files

A recent update to Bing’s Data Loss Prevention (DLP) capabilities has introduced a significant safeguard, effectively blocking Microsoft Copilot’s access to sensitive files within Office applications. This development is a critical step in enhancing enterprise data security as AI tools become more integrated into daily workflows.

The update primarily targets the protection of confidential information that users might inadvertently expose when interacting with AI assistants. By restricting Copilot’s ability to process data from sensitive Office documents, Microsoft aims to bolster the security posture for organizations leveraging these advanced AI functionalities.

Understanding the New Bing DLP Update

This enhancement to Bing’s DLP features signifies a proactive approach to managing the risks associated with AI-driven productivity tools. Previously, the integration of AI assistants like Copilot into the Microsoft 365 ecosystem offered immense productivity gains but also raised concerns about potential data exfiltration or misuse.

The core of the update lies in its ability to leverage Microsoft Purview Data Loss Prevention policies. These policies can now be specifically configured to identify and exclude sensitive files from Copilot’s processing capabilities. This means that even if a user has access to a sensitive document, Copilot will be prevented from using its content to generate responses or perform other AI-driven tasks.

The system works by detecting sensitivity labels applied to content. When a DLP policy is triggered by such a label, the identified items are effectively excluded from Copilot’s operational scope. This granular control allows organizations to strike a balance between embracing AI for efficiency and maintaining stringent data security standards.

How DLP Policies Protect Sensitive Data

Data Loss Prevention (DLP) is a crucial discipline and a set of technologies designed to detect and block the inappropriate sharing, transfer, or use of protected data. In the context of Microsoft 365, DLP policies are implemented through Microsoft Purview to identify, monitor, and protect sensitive information across various applications and devices. These policies are not merely simple text scans; they employ deep content analysis, including keyword matching, regular expressions, internal function validation, and proximity analysis to identify sensitive data. Machine learning algorithms and other methods are also utilized to ensure accurate detection.

DLP policies can act on a wide array of data transmission methods and user activities. They extend to Microsoft 365 services like Exchange, SharePoint, OneDrive, and Teams, as well as Office applications such as Word, Excel, and PowerPoint. Furthermore, Endpoint DLP capabilities monitor and control user activities involving sensitive data on Windows and macOS devices. This comprehensive coverage is vital for preventing accidental or unauthorized sharing of critical information.

When a DLP policy’s conditions are met, protective actions are taken. These actions can range from displaying a policy tip to the user, warning them about potential inappropriate sharing, to blocking the sharing entirely, with or without an override option. For instance, if a user attempts to copy a sensitive item to an unapproved location or share medical information in an email, DLP can intervene.

Copilot’s Data Handling and Security Framework

Microsoft 365 Copilot operates within the existing Microsoft 365 security and compliance framework, meaning it adheres to the data handling practices and security controls already established by an organization. A fundamental principle is that Copilot only surfaces organizational data to which individual users have at least view permissions. This ensures that Copilot does not grant access to information beyond a user’s existing authorization.

Data entered into Copilot prompts, the data it retrieves, and the generated responses remain within the Microsoft 365 service boundary, upholding existing privacy, security, and compliance commitments. Crucially, Microsoft 365 Copilot utilizes Azure OpenAI services, not OpenAI’s publicly available services, which means customer content is not cached, and modified prompts are not used for training. Prompts, responses, and data accessed through Microsoft Graph are not used to train foundation LLMs.

Microsoft Purview Information Protection also plays a role, with Copilot honoring usage rights granted to the user for encrypted data. This encryption can be applied through sensitivity labels or restricted permissions using Information Rights Management (IRM). The AI models powering Copilot are regularly updated, but these updates do not alter an organization’s security, privacy, or compliance settings. Microsoft is committed to responsible AI, guided by ethical principles, and employs a multidisciplinary team to review AI systems for potential harms and implement mitigations.

Implementing DLP Policies for Copilot

Organizations can implement Data Loss Prevention (DLP) policies specifically targeting Microsoft 365 Copilot and Copilot Chat to enhance data protection. This involves creating DLP policies within Microsoft Purview that utilize the “Microsoft 365 Copilot and Copilot Chat” policy location. These policies can be configured with conditions such as “Content contains > Sensitive information types” or “Content contains > Sensitivity labels”.

A key protective action is to block files and emails with sensitivity labels from being processed by Copilot. While identified items may still appear in citations within responses, their actual content will not be used or accessed by Copilot. This feature is generally available for Microsoft 365 Copilot, Copilot Chat, and Copilot in Word, Excel, and PowerPoint.

To set this up, administrators create DLP policies using the custom policy template and select the Microsoft 365 Copilot and Copilot Chat policy location, which disables other locations for that policy. DLP alerts, notifications, and policy simulation modes are supported, though updates to a DLP policy can take up to four hours to reflect in the Copilot experience. The primary goal is to prevent Copilot from processing sensitive prompts and to exclude files with sensitivity labels from its operations.

Key DLP Policies for Copilot Security

Microsoft Purview DLP for Copilot offers organizations the tools to balance productivity with strict data security. Three essential DLP policies are recommended to secure AI interactions: a Highly Confidential Data Exclusion Policy, a Customer Data Protection Policy, and a Legal and Compliance Document Policy. The Highly Confidential Data Exclusion Policy is paramount, preventing Copilot from processing any content labeled with the strictest sensitivity, thus isolating critical organizational data like financial records and strategic plans.

The Customer Data Protection Policy is essential for organizations handling personal data under regulations such as GDPR or CCPA. This policy targets content classified with customer data and prevents Copilot from accessing or analyzing it. The Legal and Compliance Document Policy specifically protects legal documents, contracts, and compliance-related materials due to their sensitive nature and potential attorney-client privilege considerations.

Implementing these policies provides a solid security foundation, protecting sensitive information while enabling the benefits of AI-powered productivity tools. Organizations with strong foundational DLP policies are better positioned to leverage new AI features while maintaining security and regulatory compliance.

The Role of Sensitivity Labels

Sensitivity labels are integral to the effectiveness of DLP policies when used with Copilot. These labels, applied to documents and data items, act as a signal to the DLP system, indicating the sensitivity level of the content. For example, a label such as “Confidential” can be configured to prevent Copilot from generating or using that document’s content.

When a DLP policy is created, it targets these specific sensitivity labels. If a document or data item has the designated sensitivity label, the DLP policy will then prevent Copilot from using that content for generating new information or in any of its processes. This means that data labeled with a specific sensitivity level will be effectively excluded from Copilot’s operations.

The integration of sensitivity labels with DLP policies ensures that Copilot respects the organization’s data classification and protection strategy. Sensitivity labels are deeply integrated with Copilot, providing an additional layer of protection, and their protection settings are automatically inherited by new content created from labeled files. This ensures a consistent application of security policies across an organization’s data landscape.

Broader Impact on Data Security and Compliance

The expansion of DLP controls to block Copilot’s access to sensitive Office documents, regardless of their storage location, marks a significant advancement in enterprise data security. Previously, DLP policies primarily applied to files stored in SharePoint or OneDrive, but this update extends protection to local device storage as well. This addresses customer feedback requesting more consistent protection coverage across both local and cloud-based file locations.

This enhancement means that Copilot will be unable to read or process Word, Excel, or PowerPoint documents flagged as restricted by DLP controls, irrespective of whether they are stored on a local device, in SharePoint, or OneDrive. The Office clients and the Augmentation Loop (AugLoop) component have been updated to read a file’s sensitivity label directly from the client, enabling uniform DLP enforcement. This update will be automatically enabled for organizations with existing DLP policies configured to block Copilot processing of sensitivity-labeled content, requiring no administrative action.

This development ensures consistent DLP enforcement without altering Copilot’s core capabilities, providing a more robust security posture for organizations using Microsoft 365 Copilot. It underscores Microsoft’s commitment to evolving its security measures to keep pace with the increasing integration of AI in the workplace.

User Training and Awareness

Effective employee training is paramount in navigating the complexities of AI tools and ensuring data security. Organizations must educate their workforce on company policies regarding AI usage and legal requirements such as GDPR and CCPA. This includes emphasizing the control of data input, advising employees never to submit confidential or sensitive company information to public AI platforms.

Training should also focus on monitoring AI outputs, validating all AI-generated content before sharing or acting upon it, and utilizing monitoring tools to detect potential data leaks. A culture of vigilance and openness around data safety should be fostered through ongoing training sessions. Employees need to understand that AI does not create new risks but rather exposes existing oversharing practices, making DLP a critical safety net as data governance matures.

By investing in comprehensive employee training on secure and responsible AI use, organizations can mitigate the risk of exposing sensitive content and enhance their overall security posture. This proactive approach ensures that employees are equipped to leverage AI tools effectively while upholding data protection standards.

The Future of DLP and AI Integration

The ongoing evolution of AI technologies necessitates a continuous adaptation of Data Loss Prevention strategies. As AI tools become more sophisticated and integrated into various business processes, traditional DLP solutions may prove insufficient. Modern DLP requires a comprehensive approach that extends beyond policy enforcement to include visibility into data residency, clarity on protection needs, and workflows for intelligent, rapid responses.

Microsoft Purview DLP is at the forefront of this evolution, offering unified controls that protect data across Microsoft 365 apps, endpoints, and even network communications. Future developments are expected to include enhanced network DLP capabilities and agent governance, further strengthening the security framework for AI interactions. The integration of DLP with AI is not merely a compliance exercise but a fundamental aspect of creating safe and predictable boundaries for AI operations.

Organizations that proactively implement robust DLP and DSPM controls before widely enabling AI tools gain a critical advantage in reducing the risk of AI-assisted data exposure. This strategic approach ensures that sensitive information is identified, protected, and governed by real-time enforcement rules, allowing businesses to harness the power of AI responsibly and securely.