Listen to this Post
Introduction:
The National Health Service is executing the largest healthcare AI deployment in history, rolling out Microsoft 365 Copilot to 505,000 clinicians and support staff following a trial that saved an average of 43 minutes per staff member per day. This isn’t just another productivity tool—it’s a fundamental shift in how administrative workloads are managed across the UK’s healthcare system. Microsoft Copilot Agents, built on Copilot Studio and integrated across Microsoft 365, are now automating tasks, answering queries, and retrieving information from connected knowledge sources. For NHS professionals, this means less time on documentation and more time for patient care. For IT and security teams, it means navigating a complex landscape of data protection, compliance, and AI governance.
Learning Objectives:
- Understand the technical architecture and compliance framework of Microsoft Copilot Agents within the NHS.net Connect environment.
- Master the step-by-step configuration of sensitivity labels, data loss prevention (DLP), and access controls to secure AI agents.
- Implement monitoring and governance strategies to prevent data oversharing, insider risks, and prompt injection attacks.
You Should Know:
- The NHS Copilot Ecosystem: Architecture and Compliance Prerequisites
Microsoft Copilot is available as an add-on to the Microsoft 365 suite, with Copilot Chat enabled on the NHS.net Connect shared tenant for all users. For organisations that have procured Microsoft 365 Copilot licences, users gain Generative AI capabilities across key M365 data, including SharePoint and Teams. Copilot Agents extend this functionality to interact with custom data sources and systems.
Before any NHS organisation can onboard M365 Copilot, several critical prerequisites must be met:
- Licensing: All participating users must be Enhanced Service users with M365 E3 base licences and M365 Apps for Enterprise.
- Multi-Factor Authentication (MFA): All users must have MFA enrolled.
- Device Management: Devices must be securely managed via NHS.net Connect Intune or a local solution.
- Global Sensitivity Label Policy: Organisations must opt into the Global Sensitivity Label Policy before onboarding any Copilot licences.
The Data Protection Impact Assessment (DPIA) must be reviewed and completed to demonstrate compliance with data protection laws. Additionally, the Copilot Extensibility Acceptable Use Policy (AUP) outlines rules for safe and responsible agent creation, ensuring compliance with NHS.net Connect security and data governance standards. Importantly, agents must not be used for any clinical activity, including informing clinical decision-making or direct patient care.
Linux/Windows Command for Compliance Auditing (PowerShell):
To audit MFA enrollment and device compliance across your Microsoft 365 tenant, run the following PowerShell commands:
Connect to Microsoft Graph
Connect-MgGraph -Scopes "User.Read.All", "Policy.Read.All"
Get MFA status for all users
Get-MgUser -All | ForEach-Object {
$authMethods = Get-MgUserAuthenticationMethod -UserId $<em>.Id
$mfaEnabled = $authMethods | Where-Object { $</em>.AdditionalProperties.'@odata.type' -1e 'microsoft.graph.passwordAuthenticationMethod' }
[bash]@{
User = $_.UserPrincipalName
MFAEnabled = ($mfaEnabled.Count -gt 0)
}
} | Export-Csv -Path "MFA_Audit.csv" -1oTypeInformation
- Securing Grounding Data: Sensitivity Labels and DLP Configuration
The most critical security control for Copilot Agents is protecting the “grounding data”—the information agents use to generate responses. If grounding data isn’t secured, it could lead to access beyond what users need, unauthorised disclosure, and outdated responses. Microsoft Purview provides the tools to secure agents end-to-end.
Step-by-Step Guide to Apply Sensitivity Labels:
- Opt into the Global Sensitivity Label Policy via the NHS.net Connect admin portal.
- Classify content using the predefined labels: General, Corporate, Official, and Official Sensitive.
- Block Copilot access to Official Sensitive content—all content labelled as ‘Official Sensitive’ is prevented from M365 Copilot access.
- Apply labels to emails—when sending Official Sensitive documents via email, emails must also be labelled Official Sensitive to ensure attachment content isn’t accessible.
- Use Microsoft Purview DLP to restrict sensitive files from being processed by Copilot agents.
Linux/Windows Command for Sensitivity Label Enforcement (PowerShell):
Connect to Security & Compliance Center Connect-IPPSSession Get all sensitivity labels Get-Label | Format-Table DisplayName, Identity, ContentType Apply a label to all documents in a SharePoint site Set-ComplianceTag -Identity "Official Sensitive" -RetentionAction Keep -RetentionDuration 3650
Azure CLI Command for Purview Configuration:
Create a Purview account az purview account create --1ame "NHS-Purview" --resource-group "NHS-RG" --location "uk south" Set up DLP policies az purview classification rule create --1ame "PHI-Detection" --classification-1ame "HealthcarePHI" --pattern "NHS."
- Step 2: Deploying Copilot Agents with Copilot Studio
Copilot Studio allows users (“Makers”) to design, test, and deploy AI-powered agents that automate tasks and retrieve information from connected knowledge sources. There are two types of Copilot Agents enabled for use on NHS.net Connect.
Step-by-Step Guide to Deploying a Copilot Agent:
- Acquire Licences: Each organisation must acquire Microsoft Copilot Studio User Licences and at least one Copilot Credit pack.
- Enable Environments: Only organisations on the enhanced service are provisioned with Copilot Studio licensing and environments.
- Build the Agent: Use Copilot Studio to design the agent, connect knowledge sources (SharePoint, Dataverse, external APIs), and define topics and entities.
- Test the Agent: Use the built-in test console to validate responses and identify potential prompt injection vulnerabilities.
- Publish: Publish the agent to your organisation’s environment, ensuring that MFA and device management prerequisites are met.
API Security Configuration (REST API):
Register an application in Entra ID for agent authentication
az ad app create --display-1ame "NHS-Copilot-Agent" --required-resource-access "[{'resourceAppId':'00000003-0000-0000-c000-000000000000','resourceAccess':[{'id':'e1fe6dd8-ba31-4d61-89e7-88639da4683d','type':'Scope'}]}]"
Generate a client secret
az ad app credential reset --id <app-id> --append
Get an access token
curl -X POST https://login.microsoftonline.com/<tenant-id>/oauth2/v2.0/token \
-H "Content-Type: application/x-www-form-urlencoded" \
-d "client_id=<client-id>&client_secret=<secret>&scope=https://api.botframework.com/.default&grant_type=client_credentials"
- Protecting Interactions: Enforcing Label Inheritance and Detecting Risks
Step 3 in the Microsoft Purview deployment model focuses on protecting agent interactions. This involves enforcing label inheritance in conversations and detecting risky AI interactions.
Step-by-Step Guide:
- Enforce Label Inheritance: Ensure that sensitivity labels applied to source documents are inherited by any content generated by Copilot Agents.
- Enable Microsoft Purview Insider Risk Management: Configure policies to detect unusual or risky interactions with Copilot Agents.
- Monitor Prompt Injection Attempts: Custom agents in Copilot Studio include built-in protection against user-injected prompt attacks (UPIA), but additional monitoring is recommended.
- Set Up Alerts: Configure alerts for unauthorised attempts to access sensitive data through agent interactions.
PowerShell Command for Monitoring Agent Interactions:
Search audit logs for Copilot interactions Search-UnifiedAuditLog -StartDate (Get-Date).AddDays(-7) -EndDate (Get-Date) -Operations "CopilotInteraction" | Select-Object -Property UserId, Operations, AuditData | Export-Csv -Path "Copilot_Audit.csv" -1oTypeInformation
5. Governing Interactions: Retention, Investigation, and Compliance
Step 4 focuses on retaining, investigating, and governing AI interaction data. This is essential for compliance with GDPR, HIPAA, and NHS data governance standards.
Step-by-Step Guide:
- Configure Retention Policies: Use Microsoft Purview Data Lifecycle Management to set retention periods for Copilot interaction logs.
- Enable eDiscovery: Ensure that eDiscovery capabilities are enabled to search and export Copilot interaction data for investigations.
- Set Up Compliance Manager: Use Compliance Manager to assess your organisation’s compliance posture against regulatory frameworks (HIPAA, GDPR, ISO 27001, ISO 42001).
- Regular Audits: Conduct regular audits of Copilot usage and agent configurations.
Linux/Windows Command for Log Analysis:
Use Azure CLI to query Log Analytics for Copilot logs az monitor log-analytics query -w <workspace-id> --analytics-query "Copilot_CL | where TimeGenerated > ago(7d) | summarize Count=count() by UserId, OperationName | top 10 by Count desc"
6. Mitigating Prompt Injection and AI-Specific Threats
Prompt injection attacks are an inherent vulnerability of healthcare AI agents. A successful attack could expose entire practice databases of highly sensitive information.
Mitigation Strategies:
- Input Sanitisation: Implement input validation and sanitisation for all user prompts.
- Least Privilege Access: Ensure agents only access data the user is authorised to access.
- Agent Isolation: Treat AI agents as governed solutions rather than experiments, with environment separation.
- Regular Security Assessments: Conduct regular security assessments and penetration testing of Copilot Agents.
- Use OAuth 2.0: Secure agent connections with OAuth 2.0 and manage secrets securely.
Linux Command for API Endpoint Scanning:
Use Nmap to scan for exposed agent endpoints nmap -sV -p 443 --script http-headers <agent-endpoint-ip> Use OWASP ZAP for API security testing zap-cli quick-scan --self-contained --start-options "-config api.disablekey=true" https://<agent-endpoint>/api
What Undercode Say:
- Security Is Not Optional: The NHS rollout demonstrates that AI adoption in healthcare must be paired with enterprise-grade security controls. Sensitivity labelling, DLP, and MFA are not just best practices—they are prerequisites.
- Governance Over Experimentation: Treating AI agents as governed solutions rather than experiments is critical. Organisations must define who can create, publish, and share agents, and establish minimum guardrails.
- The Human Element: Despite automation, clinical judgement remains paramount. Agents are strictly for administrative and business support, not clinical decision-making.
- Data Protection First: The DPIA and AUP frameworks ensure that data protection is baked into the deployment, not bolted on afterwards.
- Continuous Monitoring: Insider risks and prompt injection threats require ongoing monitoring, not one-time configuration.
- Scalable Architecture: The NHS model—505,000 users across 90+ organisations—proves that enterprise AI can scale securely when properly architected.
- Training Is Key: Webinars and upskilling events are essential to ensure staff understand both the capabilities and the limitations of Copilot Agents.
Prediction:
- +1 The NHS Copilot rollout will serve as a global blueprint for secure AI deployment in public healthcare, influencing policy and procurement worldwide.
- +1 AI-driven administrative savings (43 minutes per staff member daily) will translate into measurable improvements in patient wait times and staff wellbeing.
- -1 Without continuous security audits, the risk of prompt injection and data leakage will increase as more custom agents are deployed.
- -1 Organisations that delay sensitivity labelling and DLP configuration will face compliance breaches and potential regulatory fines.
- +1 The integration of Copilot Agents with Dynamics 365 and patient portals will create a seamless, secure natural language interface for patient-clinician interactions.
- -1 Insider risk will remain a persistent threat, requiring ongoing investment in Microsoft Purview Insider Risk Management and employee training.
- +1 The success of this deployment will accelerate Microsoft’s investment in healthcare-specific AI security features, including enhanced HIPAA and GDPR compliance tools.
- -1 NHS Trusts that fail to complete their own DPIAs before onboarding will face significant legal and operational risks.
- +1 By October 2026, when the rollout is completed, the NHS will have the most advanced AI-powered administrative infrastructure of any healthcare system globally.
- +1 The lessons learned from this deployment will inform AI governance frameworks across the entire public sector, setting new standards for responsible AI adoption.
▶️ Related Video (82% Match):
🎯Let’s Practice For Free:
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
IT/Security Reporter URL:
Reported By: Imranrashidhx Nhs – Hackers Feeds
Extra Hub: Undercode MoN
Basic Verification: Pass ✅
