Security Copilot Data Sharing: When Your Incident Data Helps Train a Vendor's AI, What Microsoft Confirms and the Settings to Check

Security tools now come with AI assistants that read your alerts, incident notes and investigation chats. Some of them can send a copy of that material back to the vendor to improve the vendor's AI. Microsoft's own documentation for Security Copilot says data sharing "is turned on by default" for many customers. That means the record of how your network was attacked and defended may be leaving your environment unless someone has opened one settings page.

This is not a breach, and it is not hidden. Microsoft documents it and gives administrators switches to turn it off. The risk is quieter: incident data is some of the most sensitive information a business has, and it is easy to agree to share it during a setup wizard without anyone noticing.

What Security Copilot data sharing covers

Microsoft's page "Privacy and data security in Microsoft Security Copilot" (dated April 8, 2026, updated July 13, 2026) lists what the product treats as Customer Data:

• Prompts users type into Security Copilot
• Information retrieved to generate responses
• The responses themselves
• Pinned items and file uploads

In a security tool, that means alert details, user names, device names, investigation questions and anything an analyst pastes in or uploads.

The same page lists two data sharing settings. In Microsoft's words:

1. Allow Microsoft to capture data from Security Copilot to validate product performance using human review. Microsoft says this is used to check the quality of responses, learn what tasks customers use the product for, and improve plugins and agents.
2. Allow Microsoft to capture and human review data from Security Copilot to build and validate Microsoft's security AI model. Microsoft says captured data "is used to develop security specific models built on top of Azure OpenAI foundational model or Microsoft built models."

Microsoft also says shared data is not shared with OpenAI, not used for sales, not shared with third parties, and not used to train the Azure OpenAI foundational models. Read those two statements together. Your data does not train the base model. But with the second setting on, it can help build Microsoft's own security models on top of that base model, and people at Microsoft can review it.

What changes when data sharing is on

The same Microsoft page says:

• Location: when data sharing is on, prompts and responses "may be stored outside of the tenant's chosen Workspace geo." Uploaded files stay in your workspace, but content pulled from them to answer a question can be stored somewhere else.
• Retention: the team evaluating shared data keeps it for 90 days.
• Opting out: no new data is shared after you opt out. The page gives two different figures for data that was already shared. One section says it is deleted within 30 days. Another says it is kept "for not more than 180 days." Plan around the longer figure until Microsoft clarifies.
• Feedback: user feedback on responses is collected separately. The page says an owner can only turn it off by opening a support ticket with Microsoft.

Microsoft 365 E5 and E7 tenants have different defaults

Security Copilot is included with some Microsoft 365 E5 and E7 licenses, and Microsoft sets it up automatically for eligible tenants. Microsoft's auto-provisioning page (dated June 18, 2026) says both data sharing settings start Off for those tenants. Other defaults still matter:

• Access to Microsoft 365 services data is On by default. Security Copilot can pull Microsoft Purview data, such as data loss prevention alerts, insider risk alerts and eDiscovery review sets.
• Owner access is inherited. Global Administrator, Security Administrator, Conditional Access Administrator, Intune Administrator and several compliance roles all get it automatically, and anyone in those roles can change the sharing settings later.
• Prompt location: outside the EU, prompts are processed globally (US, UK, EU or Australia and New Zealand), depending on available GPU capacity.

So "it came with E5" does not mean "nobody needs to look at it." The starting point is different, and more people can change it than most teams realize.

What is claimed, and what is not known

Confirmed: the settings, defaults, retention periods and storage rules above. They all come from Microsoft's own documentation.

Claimed: industry commentary argues that security vendors are increasingly feeding incident records into AI model development. That is plausible, and settings like Microsoft's fit it. But we found no primary source measuring how common the practice is across vendors.

Not known: what Microsoft's security models have learned from shared customer data, how human reviewers are vetted and where they work, or whether details from one customer's data could ever show up in another customer's answers. We found no public report of that happening. We also cannot tell you what your tenant chose during setup. Only your admin portal shows that.

Why incident response data privacy matters for a business

An incident record is a map of your business. It names your servers and accounts, shows which control failed and which gap is still open, and records how your team responds. A second copy at a vendor has its own retention rules, its own staff access, and its own exposure to a breach or a legal demand. You control none of those. For healthcare, legal, finance and government contractors, storing that data outside your chosen region can also conflict with contracts or what regulators expect.

The same questions apply to any vendor that runs AI on your security data: managed detection providers, SIEM platforms, endpoint protection tools and outsourced help desks.

Security Copilot data sharing checklist for this week

Questions to send to every security vendor:

• Is our data used to train or improve any AI model? Is that opt-in or opt-out, and what is the default?
• Who can review our data by hand, and where are they?
• How long is our data kept after we opt out?
• Does this also cover anything your analysts paste into AI tools on our behalf?

AI training on security data: questions people are asking

Does Microsoft Security Copilot train AI on my data?

Microsoft says your data is not used to train the Azure OpenAI foundational models. If the second data sharing setting is on, captured data is used to build and validate Microsoft's own security AI models, and people may review it.

Is Security Copilot data sharing on by default?

For customers without Microsoft 365 E5 or E7, Microsoft's documentation says yes. For E5 and E7 tenants that Microsoft sets up automatically, both sharing settings start off.

How do I turn off Security Copilot data sharing?

A Copilot owner opens the Security Copilot portal, goes to Settings, then Owner settings, and changes the data sharing selection.

How long does Microsoft keep shared Security Copilot data?

The team evaluating it keeps shared data for 90 days. For data shared before you opt out, Microsoft's page says both "within 30 days" and "not more than 180 days."

Should we stop using AI security tools?

No. Make the sharing decision on purpose, write it down, and check it again whenever your licenses change.

The bigger lesson

Every new AI feature in a security product also creates a new path for your data, and the vendor sets the defaults, not you. Treat AI settings like firewall rules: someone owns them, they are documented, and they get reviewed.

Need help reviewing your security tools and their AI settings? See security hardening or request a quote.

Get security alerts by email, with the free checklist, at security alerts.

Sources: Microsoft Learn: Privacy and data security in Microsoft Security Copilot; Microsoft Learn: Security Copilot auto provisioning for Microsoft 365 E5 and E7.