Skip to content
Aventail™ investigation into UK government AI dependence featured in Computer Weekly.Read the story
Aventail™
All guides

Is Copilot safe for work?

16 min read

Is Copilot safe for work? A dated, sourced verdict for Microsoft Copilot and GitHub Copilot by plan: training, retention, UK and EU processing, audit, prices.

It depends on which Copilot and which account: the business plans of Microsoft Copilot and GitHub Copilot are safe for most company data, and the personal plans are not. "Copilot" is two product families with separate terms, so each gets its own verdict. The short answer, as of 15 November 2026:

Microsoft Copilot (chat, Word, Excel, Outlook, Teams)

Product

Safe for company data?

Why

Copilot on a personal Microsoft account

No

Consumer terms, no admin, no audit record. Older versions of the app train on chats unless each person opts out. The free app can show ads based on chat history.

Copilot Chat on a work account

Yes, for most work

Enterprise data protection at no extra cost: Microsoft's Data Protection Addendum applies, nothing trains foundation models, and prompts are logged for audit. It only sees files a person gives it.

Microsoft 365 Copilot (paid add-on)

Yes, after a permissions clean-up

Same protection, plus access to every file, email and chat the user can already open. Loose SharePoint permissions become visible answers.

GitHub Copilot (code)

Plan

Safe for company code?

Why

Free, Pro, Pro+, Max

No

Since 24 April 2026 GitHub may train on prompts, outputs and code snippets unless each person opts out. No admin, no content exclusion, no IP indemnity.

Business

Yes, for most code

No training, prompts from the IDE are not retained, admins set policies and exclude files. The audit log does not hold prompts.

Enterprise

Yes

Same data terms as Business with more usage and models. On GitHub Enterprise Cloud, data can be held in the EU, Australia, the US or Japan. There is no UK region.

Everything below comes from Microsoft's and GitHub's own pages, read on 15 November 2026 and listed at the end. A naming point: Microsoft's documentation now calls Microsoft 365 Copilot "Microsoft Copilot" and Microsoft 365 Copilot Chat "Microsoft Copilot Chat". Licences and the pricing page still say Microsoft 365 Copilot, so this page does too. The account type decides the terms: a personal Microsoft account, or a work account in Microsoft Entra ID.

What Microsoft does with your data, by product

Personal account

Copilot Chat, work account

Microsoft 365 Copilot

Trains on your content

App updated since 18 August 2026: Microsoft says prompts, responses and files are not used to train foundation models. Older app: yes, unless the user opts out

No

No, including data read through Microsoft Graph

Terms

Microsoft Services Agreement and Privacy Statement

Data Protection Addendum and Product Terms, Microsoft as processor

Same

What it can read

What the user types or uploads

Web results, plus files the user pastes, uploads or has open. In Outlook it can read the user's mail and calendar

Everything the user has at least view permission for in SharePoint, OneDrive, Exchange and Teams

How long chats are kept

Older app: 18 months by default. Updated app: kept for the user to revisit; the pages we read give no period

Stored in a hidden folder of the user's Exchange mailbox until deleted. Admins set retention with Purview

Same

Human review

Yes, automated and human, with no opt-out (older app)

Not stated separately

Microsoft says Copilot has opted out of Azure OpenAI abuse monitoring, which includes human review

Where it is processed

Not stated per region

Nearest data centres, other regions at peak. EU traffic stays in the EU Data Boundary unless flex routing is on

Same

Web search

On; the user can turn it off

Bing queries of a few words, with no user or tenant identifiers. Outside the Data Protection Addendum and the EU Data Boundary

Same

Other model providers

Not stated

Anthropic as a Microsoft subprocessor, where licensed

Anthropic: on by default in most commercial tenants, off by default in the EU, EFTA and UK

Audit trail

None

Purview Audit (Standard), automatic, 180 days. Prompts and responses searchable with eDiscovery

Same, with the files Copilot read and their sensitivity labels

Admin controls

None. Block personal sign-in with Tenant Restrictions v2

Block or scope the Copilot app, Edge policies, data loss prevention

Sensitivity labels, DLP for Copilot, Restricted Content Discovery, SharePoint Advanced Management, agent allow-list

Standards Microsoft lists

Not listed

GDPR, ISO 27018, HIPAA support (not for web search)

GDPR, ISO 27001, ISO 42001, HIPAA

Price

Free; may show ads without a Microsoft 365 subscription

No additional cost with an eligible Microsoft 365 subscription

$30 or £23.10 per user a month, paid yearly, on top of a qualifying Microsoft 365 plan. Copilot Business add-on for Business plans: $21 or £16.10, or $18 or £13.80 for the first year if bought by 31 December 2026. UK prices exclude VAT

Commercial customers also get Microsoft's Customer Copyright Commitment, which covers copyright claims over Copilot output if the built-in filters were left on.

What GitHub does with your code, by plan

Free, Pro, Pro+, Max

Business

Enterprise

Trains on your code

Yes since 24 April 2026: inputs, outputs, code snippets and context. Opt out in Copilot settings › "Allow GitHub to use my data for AI model training" › Disabled

No. The Data Protection Agreement prohibits it

No

Retention

Not published per plan; the Privacy Statement applies

IDE, Copilot app and CLI: prompts and outputs not retained by default. GitHub.com and other surfaces: retained for the life of the account. Engagement data: two years

Same

Who runs the models

OpenAI, Anthropic, Google, xAI, Microsoft and Fireworks AI, by model. Free plan completions run on Fireworks AI

Same providers. Completions run on Azure. GitHub states zero data retention agreements with OpenAI, Anthropic and xAI

Same

Models that keep data

Claude Fable 5, 5.1 and 5.2: Anthropic retains prompts and outputs by default

Off until an admin enables them. Eligible enterprises can request zero retention until the end of 2026

Same

Where data is stored

US by default

US by default. EU, Australia, US or Japan on GitHub Enterprise Cloud with data residency

Same. No UK region

SSO and SCIM

None

Through the GitHub plan underneath

SAML or OIDC and SCIM with GitHub Enterprise Cloud

Audit trail

None for an employer

Settings, policy and licence changes, and agent activity on GitHub.com, kept 180 days, with streaming to a SIEM. No prompts

Same

Content exclusion

No

Yes, by repository, organisation or enterprise

Yes

Public code matching

User chooses Allow or Block

Admin sets it for everyone

Admin sets it for everyone

IP indemnity

Not included

Included

Included

Certifications

Not listed

SOC 2 Type 2 and ISO 27001 scope

Same

Price

$0; $10; $39; $100 a month

$19 per seat a month

$39 per seat a month

Paid plans include a monthly allowance of AI credits. Extra usage costs $0.01 a credit.

What that means for your staff

On a personal Microsoft account, there is no contract between Microsoft and your company, no admin and no log. Whether chats train models depends on which version of the app a person has, and you cannot check that.

On Copilot Chat with a work account, the training risk is gone and every prompt is logged in your tenant. A green shield in the app shows the protection.

On Microsoft 365 Copilot, the assistant reads with the user's permissions. The exposure is inside your own tenant, covered next.

On a personal GitHub Copilot plan, a developer's prompts and code snippets are training data by default. GitHub says it excludes anyone whose account is a member or outside collaborator of a paid organisation. Check that covers your contractors and any code hosted elsewhere.

On Business or Enterprise, code sent from the editor is not kept and not trained on. You get no record of prompts, because GitHub's audit log covers administration and agent activity only.

Oversharing: the Microsoft 365 Copilot risk that is yours to fix

Microsoft 365 Copilot "only surfaces organizational data to which individual users have at least view permissions". It does not widen access. It makes existing access easy to use: a salary file on a site shared with everyone is now one question away.

Microsoft's deployment blueprint puts "remediate oversharing" first. Its tools:

  • SharePoint Advanced Management, included with the Copilot licence, to find and reduce oversharing and clean up inactive sites.

  • Restricted Content Discovery, a site-level switch that keeps a SharePoint site out of Copilot answers and organisation-wide search while you review it. It does not change permissions and does not cover OneDrive.

  • Sensitivity labels and DLP. A DLP policy for the Copilot location can stop Copilot using labelled files and emails, and stop prompts that contain sensitive data types from reaching web search.

  • Restricted SharePoint Search, the older 100-site allow-list, is retiring. New enablement has been blocked since 31 July 2026.

Labels can fail. From 21 January 2026 a code error let Copilot Chat summarise emails in Sent Items and Drafts that carried a confidential label and a DLP policy. Microsoft began rolling out a fix in early February.

The EU Data Boundary and UK processing

For EU and EFTA tenants, Microsoft 365 Copilot and Copilot Chat are EU Data Boundary services: prompts are stored and processed inside it, with three exceptions.

  1. Flex routing. At peak load, inference can run in the United States, Canada or Australia. It is on by default for eligible tenants created after 25 March 2026, and an admin can turn it off.

  2. Web search. Bing queries are outside the boundary.

  3. Claude models. Anthropic models are excluded from the boundary today and off by default in the EU, EFTA and UK. On 5 November 2026 Microsoft said it would bring them inside the boundary for the EU and EFTA by the end of 2026.

The UK is outside the EU Data Boundary. A tenant provisioned in the UK has Copilot interaction content stored at rest in the UK under the Product Terms. For processing, Microsoft says customers outside the EU "may have their queries processed in the US, EU, or other regions". UK in-country inference is "expected" by the end of 2026, after an end-of-2025 target slipped. GitHub Copilot offers no UK region. Our guide to AI data residency compares this with other tools.

The red list: what never goes into any Copilot

These stay out on every plan.

  • Passwords, API keys, tokens, connection strings, private keys

  • Customer personal data: names with contact details, account numbers, health or financial records

  • Source code that is a trade secret, or anything under a customer NDA

  • Unreleased financial results, deal terms, board papers

  • Anything legally privileged

  • Security incident details while the incident is open

The risks that stay, whatever plan you buy

Prompt injection through documents and links. In June 2025 Microsoft disclosed CVE-2025-32711, a critical information disclosure flaw in Microsoft 365 Copilot. In June 2026 it patched CVE-2026-42824, which Varonis named SearchLeak: one click on a crafted link let an attacker pull emails, security codes and files with the victim's permissions. A related attack, Reprompt, hit Copilot on personal accounts. Varonis reports that enterprise customers were not affected.

Private code leaking through chat. A flaw in GitHub Copilot Chat, found in June 2025 and fixed by 14 August 2025, let hidden comments in a pull request make Copilot leak secrets and code from private repositories. In August 2025 Microsoft also patched CVE-2025-53773, a command injection in GitHub Copilot and Visual Studio that allowed local code execution.

Content exclusion has gaps. GitHub says it "is not a filesystem security boundary". It does not apply to symlinks or repositories on remote filesystems, and third-party MCP servers may ignore it.

Licence matches and wrong answers. GitHub says fewer than 1% of suggestions match public code, and the Block setting hides suggestions that match about 150 characters of it. Microsoft says Copilot responses "aren't guaranteed to be 100% factual". Someone who understands the output checks it before it ships.

Settings to change today

Microsoft 365 tenant:

  1. Tell staff to sign in to Copilot with their work account, and block personal Microsoft accounts with Tenant Restrictions v2. Microsoft advises against blocking copilot.cloud.microsoft itself.

  2. Before you assign Microsoft 365 Copilot licences, review sharing with SharePoint Advanced Management. Put HR, finance and legal sites under Restricted Content Discovery until then.

  3. Create a DLP policy for the Copilot location that blocks labelled files and emails.

  4. Set a Purview retention policy for Copilot interactions, and confirm auditing is on.

  5. EU and EFTA tenants: open Copilot › Settings › View all › Flex routing during peak load periods, and choose.

  6. Check the "AI providers operating as Microsoft subprocessors" setting. Leave "Anthropic models with Data Retention" off unless legal has accepted Anthropic's terms: Anthropic then stores most inputs and outputs for up to 30 days, outside your Microsoft agreement.

GitHub organisation:

  1. Give every developer a Business or Enterprise seat. Their individual plan is cancelled when they join.

  2. Anyone still on a personal plan: Copilot settings › "Allow GitHub to use my data for AI model training" › Disabled.

  3. Set "Suggestions matching public code" to Block.

  4. Add content exclusion for secrets, configuration and customer data paths.

  5. Leave the Claude Fable models disabled unless your enterprise has the zero-retention exemption.

  6. Stream the audit log to your SIEM. GitHub keeps 180 days.

What to put in the acceptable use policy

Name the approved products exactly: Copilot Chat or Microsoft 365 Copilot on a work account, GitHub Copilot Business or Enterprise. Ban personal accounts and personal plans for work. Add the red list and where to report a mistake without blame. Our guide to shadow AI and acceptable use policies has a policy outline.

Keep sensitive work on approved models

Microsoft Copilot sends prompts to models that Microsoft or its subprocessors host. GitHub says Copilot Business and Enterprise need a live connection to its proxy service and do not work in air-gapped environments. GitHub's docs also describe bring-your-own-key models in some clients. For work on the red list, use a tool whose requests you can keep on approved machines.

Aventail™ Connect configures supported AI apps to send their model requests through Aventail™ Link, an inference engine and gateway that runs on each machine, under routing rules set in Aventail Control. Claude Desktop, Claude Code, OpenCode and Codex are supported today. Neither Copilot is on that list. A rule can say that a team's requests run on the device only. When no allowed destination exists, the request fails with an error and is not sent elsewhere. The hosted Control service receives management, device and routing events, and does not receive prompts or responses.

That gives you two lanes:

  • Copilot on a work account for general work and non-restricted code, with the settings above.

  • A coding assistant routed by Aventail™ for anything on the red list: a local model on the developer's machine, or an approved provider through your own account, with request records in Control that show where each request ran.

Setup is the Connect button in Link, and the free plan covers three connected devices. The steps are in our guide to Claude Code with Aventail™ Connect.

Frequently asked questions

Is Microsoft Copilot safe for confidential information?
On a work account, yes for most of it: prompts fall under Microsoft's Data Protection Addendum and are not used for training. The red list stays out, and Microsoft 365 Copilot needs a permissions review first. On a personal account, no.

Does Microsoft Copilot use my data for training?
Not on a work account. On a personal account it depends on the app version: Microsoft says the app updated since 18 August 2026 does not train foundation models on prompts and files, and the older app does unless the user opts out.

Can my employer see my Copilot chats?
Yes, on a work account. Prompts and responses are logged, and admins can search them with Microsoft's audit and eDiscovery tools.

Is GitHub Copilot safe for proprietary code?
On Business or Enterprise, yes for most code: GitHub does not train on it and does not retain prompts sent from the IDE. On Free, Pro, Pro+ or Max, training is on by default.

Can Microsoft 365 Copilot read all my company's files?
It reads what each user already has permission to view. In a tenant with loose sharing, that is more than people expect.

Where does Copilot process UK data?
Microsoft stores Copilot content for UK tenants in the UK. Inference can run in other regions until UK in-country processing arrives, expected by the end of 2026. GitHub has no UK region.

Is the free version of Copilot safe for work?
On a personal account, no. Copilot Chat on a work account costs nothing extra with an eligible Microsoft 365 subscription and carries enterprise data protection.

Sources

All read on 15 November 2026.

  1. Microsoft Learn, Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat: https://learn.microsoft.com/en-us/copilot/microsoft-365/enterprise-data-protection

  2. Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy

  3. Microsoft Learn, Copilot Chat privacy and protections: https://learn.microsoft.com/en-us/copilot/privacy-and-protections

  4. Microsoft Learn, Microsoft Copilot overview and licences: https://learn.microsoft.com/en-us/copilot/overview

  5. Microsoft Learn, Manage Copilot Chat: https://learn.microsoft.com/en-us/copilot/manage

  6. Microsoft Support, Privacy FAQ for Microsoft Copilot (older app): https://support.microsoft.com/en-us/topic/privacy-faq-for-microsoft-copilot-27b3a435-8dc9-4b55-9a4b-58eeb9647a7f

  7. Microsoft Support, Microsoft Copilot for individuals: your data, privacy, and responsible AI: https://support.microsoft.com/en-us/privacy/microsoft-copilot/overview

  8. Microsoft Support, Microsoft Copilot for individuals: your activity history: https://support.microsoft.com/en-us/privacy/microsoft-copilot/activity-history

  9. Microsoft Support, Microsoft Copilot for individuals: your privacy controls and choices: https://support.microsoft.com/en-us/privacy/microsoft-copilot/privacy-controls

  10. Microsoft Support, Data protection when using Microsoft Copilot Chat for work or school: https://support.microsoft.com/en-us/privacy/data-protection-when-using-microsoft-365-copilot-chat-for-work-or-school

  11. Microsoft Learn, Flex routing: https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-flex-routing

  12. Microsoft Learn, Claude models from Anthropic in Microsoft Online Services: https://learn.microsoft.com/en-us/copilot/microsoft-365/connect-to-ai-subprocessor

  13. Microsoft, Bringing Claude models in Microsoft Copilot into the EU Data Boundary (5 November 2026): https://www.microsoft.com/en-us/copilot/blog/2026/11/05/bringing-claude-models-in-microsoft-copilot-into-the-eu-data-boundary/

  14. Microsoft, In-country data processing for Microsoft 365 Copilot (4 November 2025, updated 3 April 2026): https://www.microsoft.com/en-us/microsoft-365/blog/2025/11/04/microsoft-offers-in-country-data-processing-to-15-countries-to-strengthen-sovereign-controls-for-microsoft-365-copilot/

  15. Microsoft Learn, Data residency for Microsoft 365 Copilot: https://learn.microsoft.com/en-us/microsoft-365/enterprise/m365-dr-workload-copilot

  16. Microsoft Learn, Product Terms data residency: https://learn.microsoft.com/en-us/microsoft-365/enterprise/m365-dr-product-terms-dr

  17. Microsoft Learn, Audit logs for Copilot and AI applications: https://learn.microsoft.com/en-us/purview/audit-copilot

  18. Microsoft Learn, Microsoft Purview auditing solutions: https://learn.microsoft.com/en-us/purview/audit-solutions-overview

  19. Microsoft Learn, Retention for Copilot and AI apps: https://learn.microsoft.com/en-us/purview/retention-policies-copilot

  20. Microsoft Learn, Microsoft Purview protections for generative AI apps: https://learn.microsoft.com/en-us/purview/ai-microsoft-purview

  21. Microsoft Learn, Data Loss Prevention for Microsoft 365 Copilot, Copilot Chat and Cowork: https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about

  22. Microsoft Learn, Restricted SharePoint Search: https://learn.microsoft.com/en-us/sharepoint/restricted-sharepoint-search

  23. Microsoft Learn, Restrict discovery of SharePoint sites and content: https://learn.microsoft.com/en-us/sharepoint/restricted-content-discovery

  24. Microsoft Learn, Secure and govern Microsoft Copilot blueprint: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-blueprint-oversharing

  25. Microsoft, Copilot plans for enterprise (US and UK pages): https://www.microsoft.com/en-us/copilot/pricing/enterprise and https://www.microsoft.com/en-gb/copilot/pricing/enterprise

  26. Microsoft, Copilot plans for business (US and UK pages): https://www.microsoft.com/en-us/copilot/pricing/business and https://www.microsoft.com/en-gb/copilot/pricing/business

  27. GitHub Docs, Plans for GitHub Copilot: https://docs.github.com/en/copilot/get-started/plans

  28. GitHub, Copilot plans and FAQ: https://github.com/features/copilot/plans

  29. GitHub Copilot Trust Center, FAQ: https://copilot.github.trust.page/faq

  30. GitHub Docs, Managing GitHub Copilot policies as an individual subscriber: https://docs.github.com/en/copilot/how-tos/manage-your-account/manage-policies

  31. GitHub Docs, Hosting of models for GitHub Copilot: https://docs.github.com/en/copilot/reference/ai-models/model-hosting

  32. GitHub Docs, Content exclusion for GitHub Copilot: https://docs.github.com/en/copilot/concepts/context/content-exclusion

  33. GitHub Docs, References to matching public code: https://docs.github.com/en/copilot/concepts/completions/code-referencing

  34. GitHub Docs, Reviewing audit logs for GitHub Copilot: https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-organization/review-activity/review-audit-logs

  35. GitHub Docs, GitHub Copilot policies for enterprises and organizations: https://docs.github.com/en/copilot/concepts/policies

  36. GitHub Docs, About GitHub Enterprise Cloud with data residency: https://docs.github.com/en/enterprise-cloud@latest/admin/data-residency/about-github-enterprise-cloud-with-data-residency

  37. CVE record, CVE-2025-32711 (published 11 June 2025): https://www.cve.org/CVERecord?id=CVE-2025-32711

  38. CVE record, CVE-2026-42824 (published 4 June 2026): https://www.cve.org/CVERecord?id=CVE-2026-42824

  39. Varonis, SearchLeak (15 June 2026): https://www.varonis.com/blog/searchleak

  40. Varonis, Reprompt (page dated 16 June 2026): https://www.varonis.com/blog/reprompt

  41. Petri, Microsoft 365 Copilot bug exposes confidential emails despite DLP safeguards (19 February 2026): https://petri.com/microsoft-365-copilot-bug-confidential-emails/

  42. Legit Security, CamoLeak: https://www.legitsecurity.com/blog/camoleak-critical-github-copilot-vulnerability-leaks-private-source-code

  43. CVE record, CVE-2025-53773 (published 12 August 2025): https://www.cve.org/CVERecord?id=CVE-2025-53773

  44. Aventail™, Connect overview: https://docs.aventail.co.uk/docs/connect/overview

  45. Aventail™, Trust: https://aventail.co.uk/trust

We use optional cookies and tracking to identify visiting companies, load our contact form, measure Google Ads download clicks and understand page use with PostHog on our homepage, Connect and How it works pages, Ollama and LM Studio comparison guides and campaign pages. Cookie policy