When your staff paste customer lists, financial forecasts or source code into ChatGPT, that data leaves your control and sits on OpenAI's servers. The same risk applies to Claude, Gemini and hundreds of other public AI tools. You cannot rely on policy documents alone. This guide shows the technical controls, DLP rules and policy steps that prevent employees pasting data into ChatGPT and other public AI endpoints, while still letting teams use approved AI tools for legitimate work.
In this guide:
- Why ChatGPT and similar tools are a data leakage risk
- What data has already left through public AI tools
- Network, browser and DLP controls that block unauthorised AI endpoints
- How to offer safe alternatives so staff can still benefit from AI
- Policy and training that reinforce the technical controls
- Microsoft 365 Copilot and Azure OpenAI as controlled replacements
Why public AI tools create a data leakage risk
ChatGPT, Claude, Gemini and similar tools are designed to remember conversation history. When an employee pastes a customer email, a pricing spreadsheet or a draft contract into the prompt box, that text travels to the vendor's cloud and is stored. The free tier of ChatGPT explicitly uses conversation data to train future models. Even paid enterprise accounts retain prompts for abuse monitoring and service improvement unless you negotiate specific contract terms.
The risk is not theoretical. In 2023 Samsung restricted staff use of ChatGPT after engineers pasted internal source code and meeting notes into it. The same pattern appears wherever staff handle client details, budget files or candidate CVs.
Public AI endpoints also sit outside your security perimeter. You have no audit trail of what was sent, no way to recall the data and no visibility into who else might retrieve similar prompts through model outputs. If the account uses a personal email address, you cannot even see usage logs.
What data has already left through shadow AI
Shadow AI is any AI tool or service that staff use without IT approval. It includes free ChatGPT accounts, browser extensions that inject AI features, and third-party services that offer "AI-powered" document summary or translation.
Most organisations discover shadow AI only after a data incident or during a security audit. Typical leakage includes:
- Customer names, email addresses and phone numbers pasted for mail-merge templates or list cleaning.
- Financial data: revenue figures, margin calculations, budget forecasts used to generate charts or summaries.
- Source code snippets sent to debug errors or convert between programming languages.
- Legal contract drafts, NDA text and terms uploaded for redlining or tone adjustment.
- Internal strategy documents, board meeting notes and project plans summarised for executive briefings.
- Personal data of employees or candidates: CVs, appraisal notes, salary bands.
You can run a shadow AI discovery scan to identify which public AI endpoints your users have already contacted and what volume of data was sent. The scan examines proxy logs, firewall traffic and browser history (where permitted) to build a list of active shadow AI tools.
Network controls: blocking AI endpoints at the firewall and proxy

The simplest technical control is to block public AI domains at your firewall or web proxy. This stops browsers and applications from reaching ChatGPT, Claude, Gemini and similar services before any data is sent.
Firewall rules
Create outbound firewall rules that deny traffic to known AI endpoint domains. The list includes:
- chat.openai.com and api.openai.com (ChatGPT and OpenAI API)
- claude.ai and api.anthropic.com (Claude)
- gemini.google.com and bard.google.com (Google Gemini)
- huggingface.co and replicate.com (model hosting platforms)
- Any third-party summarisation, translation or writing tools your scan identified
Maintain the block list in a threat intelligence feed so new AI domains are added as they appear. This approach works best for organisations with a corporate network; remote workers on home broadband will bypass the firewall unless they connect via VPN.
Web proxy and content filtering
If you run a web proxy (Zscaler, Cisco Umbrella, Palo Alto Prisma Access or similar), configure category-based filtering to block "AI and ML services" or "Generative AI" categories. The proxy vendor updates these categories automatically as new tools launch.
You can also set up an allow list: permit only approved AI endpoints (such as your own Azure OpenAI instance) and block everything else. This inverts the problem but gives tighter control.
Browser controls: Microsoft Edge and Chrome policies
Network blocks are effective for desktop traffic but do not cover mobile devices on cellular data or staff using personal laptops. Browser-level policies add a second layer.
Microsoft Edge administrative templates
Deploy Edge via Intune or Group Policy with the following settings:
- URLBlocklist: add chat.openai.com, claude.ai, gemini.google.com and other public AI domains.
- Disable Edge Copilot sidebar if you do not want users to access Bing Chat (which is powered by GPT-4) unless you have a commercial data protection agreement in place.
- Restrict installation of browser extensions: block extensions from unknown publishers to prevent AI-powered writing assistants or summarisers.
These policies travel with the user's Azure AD / Entra ID profile, so they apply even on BYOD devices enrolled in Intune.
Google Chrome enterprise policy
If your organisation uses Chrome, set the URLBlocklist policy via Google Workspace admin console or Windows Group Policy. Add the same AI endpoint domains. Also set ExtensionInstallBlocklist to prevent unapproved extensions.
Microsoft Purview DLP: detecting and blocking AI-bound sensitive data

Data loss prevention (DLP) in Microsoft Purview inspects content as it leaves your environment and blocks or alerts on policy violations. This catches attempts to paste data into AI tools even if the user evades network or browser blocks (for example, by tethering to a mobile hotspot).
Create a DLP policy for browser uploads
In the Microsoft Purview compliance portal (compliance.microsoft.com):
- Go to Data loss prevention > Policies > Create policy.
- Choose Custom template.
- Name: Block sensitive data upload to AI endpoints.
- Locations: select Devices (this requires Endpoint DLP, which needs Microsoft 365 E5 or Compliance add-on).
- Conditions: Content contains > Sensitive info types. Add credit card numbers, national IDs, customer lists (custom sensitive info type), financial data classifiers.
- Actions: when content is uploaded to a restricted cloud service domain, block the activity and send user notification.
- Restricted domains: add chat.openai.com, claude.ai, gemini.google.com.
This policy inspects clipboard and file upload operations in Edge and Chrome (when the browser is managed). If a user copies a table of customer emails and tries to paste it into ChatGPT, the paste is blocked and a notification appears.
Extend DLP to email and file shares
Users sometimes export data to a personal email account then access it from an unmanaged device to bypass desktop controls. Create a second DLP policy for Exchange and SharePoint:
- Locations: Exchange email, SharePoint sites, OneDrive accounts.
- Conditions: content contains financial data, customer records or any custom-defined sensitive information type.
- Actions: block sharing with external recipients or flag for review.
This catches side-channel leakage. For a full configuration checklist, see AI data loss prevention.
Conditional Access: restricting unmanaged devices
Microsoft Entra ID (formerly Azure AD) Conditional Access policies let you require device compliance or block access to Microsoft 365 from unmanaged devices. This reduces the risk of staff using a home laptop or a café computer to work around your controls.
Create a Conditional Access policy:
- Assignments: select all users (or the group that handles sensitive data).
- Cloud apps: select Office 365 or specific apps like SharePoint, Exchange.
- Conditions > Device platforms: all platforms.
- Grant: require device to be marked as compliant (which means Intune-enrolled and passing your compliance rules), or require hybrid Azure AD join.
An unmanaged device cannot sign in to Microsoft 365, so the user cannot retrieve company data to paste into a public AI tool. This is strict; if you need a less disruptive option, apply the policy only to devices outside your corporate network or only to high-risk roles.
Offering a safe alternative: private AI workspace
Blocking public AI tools without offering a replacement frustrates staff and encourages creative workarounds. The solution is to deploy a private AI workspace where employees can chat with GPT-4, Claude or other models without data leaving your Azure tenant.
A private AI workspace is a web application hosted in your own Azure subscription. It connects to Azure OpenAI Service (which offers GPT-4, GPT-4 Turbo and GPT-4o under a Microsoft data processing agreement) or to models you deploy yourself. All prompts and responses stay in your tenant; no data is sent to OpenAI's public cloud, and your content is never used to train any model.
How to set up a private AI chat interface
- Provision Azure OpenAI Service in your subscription (available in most regions; check current availability).
- Deploy the gpt-4 or gpt-4o model in the Azure OpenAI studio.
- Build or deploy a simple chat UI (a React or Blazor app) that calls the Azure OpenAI endpoint via REST or the .NET SDK.
- Host the UI as an Azure App Service or Static Web App, secured with Entra ID authentication (only your employees can sign in).
- Lock down the OpenAI endpoint with a private endpoint or service firewall rules so it only accepts traffic from the UI's subnet.
Users bookmark the internal chat URL and use it just like ChatGPT. You control retention: configure Azure Monitor to log prompts for audit purposes, or discard them after each session. We help build and configure these environments as part of our AI production engineering service.
Microsoft 365 Copilot: a managed, compliant alternative
Microsoft 365 Copilot is built into Word, Excel, Outlook, Teams and other apps. It uses GPT-4 but runs entirely within your Microsoft 365 tenant. User prompts and company data never train any external model, and all activity is logged in the Microsoft 365 audit log.
Copilot respects your existing permissions: a user cannot ask Copilot to summarise a file they do not have access to. It honours DLP policies, retention labels and sensitivity labels. If a document is marked confidential, Copilot will not surface that content in responses to users outside the authorised group.
Deploying Copilot removes the temptation to use public AI tools because staff have a legitimate, integrated option. Before rollout, run a Microsoft 365 Copilot readiness check to confirm your permissions model is clean (no overshared files or broad "Everyone" groups) and your data classification is accurate.
AI acceptable use policy and training
Technical controls are the foundation, but policy and training ensure staff understand why the controls exist and what they should do instead.
Publish an AI acceptable use policy
An acceptable use policy (AUP) defines which AI tools are approved, what data may be shared, and the consequences of policy breaches. The policy should:
- List approved tools: internal chat UI, Microsoft 365 Copilot, specific Azure OpenAI apps.
- List prohibited tools: ChatGPT, Claude, Gemini, any third-party summarisation or translation service not on the approved list.
- Define what constitutes company data: customer names, email addresses, financial figures, source code, strategy documents, personal data of staff or customers.
- Explain the risk: data sent to public AI tools may be retained, used for training or disclosed in future model outputs.
- Direct users to the approved alternative and explain how to access it.
- State that DLP and audit logs are active and that breaches will trigger a review.
We publish a free AI acceptable use policy template you can adapt to your organisation. Have legal review it, then distribute via email, intranet and onboarding packs.
Run a one-hour AI security workshop
Policy documents are rarely read. A short, practical training session is more effective. Cover:
- Live demo: show ChatGPT retaining a prompt and explain that OpenAI's servers now hold that data.
- Real incident example: mention the Samsung code leak or a similar public case study.
- Show the approved alternative: walk through signing in to your private AI workspace or using Copilot in Word.
- Q&A: answer questions about what's allowed (generic questions, public information) and what's not (customer data, financials).
Record the session and add it to your learning management system (LMS) for new joiners. Repeat annually or whenever you add a new AI tool to the approved list.
Monitoring and audit: catching violations before they escalate
Even with technical controls and policy in place, some users will test the boundaries. Monitoring lets you spot violations early.
Microsoft Purview audit log
Enable unified audit logging in the Microsoft Purview compliance portal. This records DLP policy triggers, file access, external sharing and device activity. Search the audit log for:
- DLP policy match: filters for blocked copy/paste events to AI endpoints.
- File accessed: unusual volume of file opens (may indicate bulk export).
- External sharing: files shared to personal email accounts.
Export the results to Excel or Power BI and review weekly. If the same user triggers multiple DLP blocks, follow up with their manager and offer a refresher on the AI policy.
Endpoint DLP activity reports
In the Purview portal, go to Data loss prevention > Reports > DLP incidents. This dashboard shows how many times sensitive content was blocked, which domains were targeted and which users were involved. Use it to identify:
- High-risk users: frequent DLP blocks suggest either legitimate workflow friction or deliberate attempts to bypass controls.
- New AI endpoints: if users attempt to reach a domain not yet in your block list, add it.
- Data classification gaps: if DLP is not triggering on data you consider sensitive, your sensitive info types or trainable classifiers need tuning.
Securing AI agents and automation workflows
AI is not only used by humans typing into chat boxes. Many teams build Power Automate flows, Azure Logic Apps or custom automation that call AI APIs. These workflows can also leak data if they send company information to public endpoints.
Inventory and audit automation
Run a quarterly audit of Power Automate flows and Logic Apps. Check for:
- HTTP actions that POST to chat.openai.com or other public AI APIs.
- AI Builder actions (Power Platform's built-in AI) that process sensitive documents: ensure these stay inside the Power Platform environment.
- Third-party connectors: some workflow connectors wrap AI summarisation or sentiment analysis; verify where the data goes.
If a flow must use AI, replace public API calls with calls to your Azure OpenAI endpoint or deploy a custom AI agent that runs in your tenant. We build these as part of AI agent development projects.
Governance layer: centralised AI usage policy

As your organisation adopts AI in more places (Copilot, private chat, Power Platform AI Builder, custom Azure OpenAI apps), you need a governance framework to maintain consistent security and compliance.
An AI governance programme includes:
- AI inventory: a register of every AI system in production, its data sources, model provider and risk classification.
- Risk assessment: score each system by data sensitivity, user base and potential compliance impact.
- Approval workflow: new AI projects (a Power Automate flow with GPT, a custom agent, a third-party summarisation API) must pass a lightweight review before deployment.
- Regular review: quarterly meetings to update the risk register, review audit reports and add new endpoints to the block list.
Governance does not need to slow innovation. A simple checklist (what data does it touch? where does it run? who can access it?) and a 30-minute review call with IT and legal are usually enough.
Industry-specific considerations
Financial services
Firms handling customer investments, payment data or credit information face strict record-keeping and data residency rules. Public AI tools break these because you lose the audit trail and cannot prove data was not disclosed. Deploy Azure OpenAI in your tenant and retain all prompts in Azure Monitor. See AI data protection for financial services for a full control set.
Software companies
Engineering teams are heavy AI users. Developers paste code into ChatGPT to debug errors, generate unit tests or convert syntax. The risk is leaking proprietary algorithms, API keys embedded in code or customer data in test fixtures. Block public endpoints and deploy GitHub Copilot for Business or Azure OpenAI with code-focused prompts. Details at AI data protection for software companies.
Manufacturing and supply chain
Bills of material, supplier contracts and production schedules are trade secrets. If your procurement team uses AI to summarise RFQs or compare supplier bids, ensure it runs on Azure OpenAI or your private workspace. Manufacturing also involves operational technology (OT); if AI is used to analyse sensor data or production logs, segment that environment from the internet entirely. More at AI data protection for manufacturing.
Compliance and regulatory alignment
Preventing data leakage to AI tools helps you meet several compliance obligations:
- GDPR (EU): sending personal data to a public AI endpoint outside your control may breach Article 28 (processor obligations) or Article 32 (security of processing).
- ISO 27001: requires organisations to control external data sharing and log access to information assets; public AI bypasses both.
- SOC 2 Type II: controls around data retention, encryption in transit and logical access; public AI tools fail all three unless you have a specific BAA or similar agreement.
- Sector-specific rules: data protection regulations and similar frameworks require consent, audit and local storage that public AI tools cannot guarantee.
You do not need to block AI entirely to remain compliant. Using Azure OpenAI with a data processing agreement or Microsoft 365 Copilot covered by your Microsoft licensing terms satisfies these requirements.
Summary: a layered defence
No single control will prevent employees pasting data into ChatGPT. A robust approach combines:
- Network and browser blocks to close the route to public AI endpoints.
- DLP policies to detect and block sensitive content in transit.
- Conditional Access to keep company data on managed devices.
- A private AI workspace or Microsoft 365 Copilot so staff have a legitimate, safe alternative.
- Policy and training to explain the why and the how.
- Monitoring and governance to catch violations and adapt controls as new tools appear.
Start with the quick wins: add public AI domains to your firewall block list and publish an acceptable use policy. Then layer on DLP, deploy a private workspace and schedule quarterly governance reviews. The goal is not to eliminate AI use, but to channel it into systems you control and audit.
Frequently asked questions
Can I detect if employees have already pasted data into ChatGPT?
You cannot retrieve past ChatGPT prompts unless you control the account. Run a shadow AI discovery scan of proxy logs, firewall traffic and browser history (where permitted by policy) to identify which users contacted public AI endpoints and estimate data volume. This tells you the scale of risk but not the exact content. For future monitoring, enable Microsoft Purview audit logging and Endpoint DLP to capture all attempted uploads to AI domains.
Does blocking ChatGPT stop all AI data leaks?
No. Blocking chat.openai.com stops direct access to ChatGPT, but hundreds of other AI tools exist: Claude, Gemini, Perplexity, niche summarisation services and browser extensions. Use category-based web filtering or maintain a block list updated by threat intelligence. Also deploy DLP to catch data leaving via any HTTP upload, regardless of destination.
Will Microsoft 365 Copilot send my data to OpenAI?
No. Microsoft 365 Copilot runs in your tenant. User prompts and company documents are processed by Microsoft's Azure OpenAI Service under your existing Microsoft data processing terms. Your content is not used to train any public model and does not leave the Microsoft cloud. Copilot respects your DLP policies, sensitivity labels and permissions.
What if I need to let one team use ChatGPT for research?
Create an exception for a specific user group. In your firewall or DLP policy, allow that group to reach chat.openai.com but require them to use a corporate ChatGPT Team or Enterprise account (which offers better data controls and audit logs). Log all usage and review it quarterly. Alternatively, give them access to your private AI workspace with a note that it is equivalent to ChatGPT but safer.
How do I handle mobile devices and remote workers?
Mobile devices on cellular data bypass your corporate firewall. Enrol them in Intune and deploy browser URL block lists via mobile app management (MAM) policies. For remote workers, require VPN or use Conditional Access to block access to Microsoft 365 from unmanaged devices. Endpoint DLP on Windows and macOS laptops will block copy/paste to AI sites even when off-network.
What is the difference between Azure OpenAI and OpenAI's ChatGPT?
Azure OpenAI Service is Microsoft's enterprise offering. It provides the same GPT-4 and GPT-4o models as ChatGPT but hosted in your own Azure subscription. All data stays in your tenant, Microsoft does not use your content for training and you control retention and logging. ChatGPT (chat.openai.com) is OpenAI's public consumer service; prompts are stored on OpenAI's infrastructure and may be used to improve models unless you have a specific enterprise agreement.
Can I monitor AI usage without blocking it?
Yes. Set DLP policies to Audit mode instead of Block. This logs every attempt to upload sensitive content to AI endpoints but does not stop the user. Use this during a pilot phase to understand usage patterns and false-positive rates. After tuning the policy, switch to Block. You can also use web proxy reporting to see which AI domains are accessed most and by whom.
Do I need Microsoft 365 E5 to run these controls?
Endpoint DLP (which blocks copy/paste to web apps) requires Microsoft 365 E5 or the Compliance add-on (E3 + standalone DLP license). Basic DLP on Exchange and SharePoint is included in E3. Conditional Access is in all Azure AD Premium P1 plans, which are included in Microsoft 365 Business Premium and E3 upwards. If you are on E3, you can deploy network blocks, browser policies and a private AI workspace without upgrading, then add Endpoint DLP when budget allows.
Need help locking down AI endpoints and deploying a private workspace for your team? We build the technical controls and policy framework in days.
See AI data loss prevention



