Shadow AI: Your Employees' Secret ChatGPT Use Is a Data Leak

Glowing documents pulled through a violet keyhole shaped portal as a blue shield approaches

The security concerns with ChatGPT come down to five things: whether your prompts are used to train the model, how long they are kept, who can open a shared link, what browser extensions and unofficial apps can read, and which company systems a connected app can reach. On a personal account most of those are decided by the employee's default settings; in a company workspace (ChatGPT Business, formerly ChatGPT Team, or ChatGPT Enterprise) OpenAI does not train on your data by default, admins control retention and app access, shared links stay inside the workspace, and sign-in runs through your identity provider. Shadow AI is what happens when staff use the personal version instead, and it has quietly become one of the most common ways confidential business data walks out the door.

Most of your employees already use AI at work. The 2024 Microsoft and LinkedIn Work Trend Index put it at 75% of knowledge workers, and found 78% of those users bringing their own AI tools to work rather than ones their company provided. This post explains what that looks like, answers the security questions owners ask about ChatGPT, and shows how a sanctioned workspace plus a one-page policy fixes most of it. For the step-by-step admin settings, see our ChatGPT Business security settings checklist; this article stays on the risk and the remedy.

What does shadow AI look like?

It is rarely malicious. It is a salesperson pasting a client contract into a free chatbot for a summary. A developer feeding error logs, credentials included, to a coding assistant. An analyst uploading the customer spreadsheet to clean it up.

Every one of those prompts is a data transfer to a third party, made with no contract, no review and no record. And adoption keeps climbing. Survey after survey shows the same shape, with most users bringing their own tools rather than waiting for permission:

Employees using unsanctioned AI tools at work (% of knowledge workers)
025507510095520222023 H12023 H22024 H12024 H2ChatGPT reaches 100M users025507510095520222023 H12024 H12024 H2ChatGPT reaches 100M users

Blended from published workplace-AI surveys; 2024 H2 reflects data to date. Industry averages; illustrative.

View data table
CategoryUnsanctioned AI use (% of knowledge workers)
20229
2023 H122
2023 H234
2024 H147
2024 H255

What are the security concerns with ChatGPT?

The concerns are real, but almost all of them change depending on which ChatGPT your staff are signed into. This table compares a personal account (Free, Go, Plus or Pro) with a company workspace, using OpenAI's own documentation as of October 2026:

ConcernPersonal account (Free, Go, Plus, Pro)ChatGPT Business or Enterprise workspace
Training on your promptsOn unless the user turns off "Improve the model for everyone"; a thumbs-up or thumbs-down can send the whole conversation for training even after opting outNot used for training by default, including GPT conversations, projects and data reached through apps
RetentionKept in the user's history until they delete it; deleted chats are removed within 30 daysWorkspace admins control retention; deleted conversations are removed within 30 days unless the law requires otherwise
Shared linksAnyone with the link can open itBusiness links open only for members of the same workspace; turning link sharing off workspace-wide is an Enterprise control
Browser extensions and unofficial appsThird-party extensions can read page content, including chats, under terms OpenAI does not controlSame risk; the fix is an extension allowlist pushed through Chrome or Edge policy
Connected apps reaching company systemsThe employee connects their own Google Drive, Outlook or Slack with no oversightApps are enabled by default in Business and admins set workspace-wide availability and per-app approval (always ask, read only, low-risk or all actions); ChatGPT uses the user's existing permissions
Who can see chatsOnly the user, and OpenAI under its policiesBusiness: private chats are not automatically visible to admins, though owners can export a member's data; Enterprise: an audit log through the Compliance Platform
IdentityAny email and passwordSAML or OIDC single sign-on and domain verification on Business; SCIM and role-based access on Enterprise; multi-factor authentication is enforced in your identity provider, not in ChatGPT

Two details that trip people up. First, OpenAI's enterprise privacy page says workspace admins can view, access, export and delete end-user conversations, while its Business help center says admins and owners cannot see private member chats by default; tell staff exactly which applies in your workspace. Second, the ChatGPT vocabulary changed in 2026: what used to be called connectors are now apps and plugins, synced connectors were retired, and access to Google Drive, SharePoint and Teams is real-time through apps. Our AI data privacy FAQ covers how the other major tools handle the same questions.

Is ChatGPT Enterprise secure?

Secure enough for most businesses, provided you configure it. OpenAI says ChatGPT Business and ChatGPT Enterprise have each completed a SOC 2 Type 2 audit, encrypt data at rest with AES-256 and in transit with TLS 1.2 or higher, and do not train on workspace data by default. Enterprise adds what Business lacks: SCIM provisioning, role-based access controls, the Compliance Platform that feeds eDiscovery, DLP and SIEM tools (Microsoft Purview among them), IP allowlisting, customer-managed encryption keys, data residency (listed for Enterprise only on OpenAI's pricing comparison, though OpenAI says it is rolling out gradually to Business at checkout) and ISO 27001, 27017, 27018 and 27701 certification. Neither plan secures what you control: who gets a seat, which apps may reach which systems, what staff paste in, and whether anyone checks the output. A secure platform with a sloppy SharePoint behind it still leaks.

Why is shadow AI a data leak, not a quirk?

The same report adds a detail that should worry you more: 52% of the people using AI at work are reluctant to admit using it for their most important tasks. Usage you cannot see is risk you cannot manage, and the risk is concrete:

  • Retention and training. Consumer AI tools may keep prompts and, depending on settings, use them to train models. Your data can end up embedded somewhere you can never audit or delete.
  • Regulated data. Client personal information or health and financial records pasted into a consumer tool have left your controlled environment with no agreement in place; in some industries that is a reportable event. Patient information in particular belongs only in a tool covered by a business associate agreement, as our guide to ChatGPT, Claude and HIPAA explains.
  • No provenance. When AI-drafted text lands in contracts and client work, nobody can say where a clause came from.
  • Cover for exfiltration. Heavy personal-account AI use looks identical to someone quietly feeding your data out on purpose.

For the broader defenses against exactly this kind of quiet exit, see our FAQ on data loss prevention for small business.

Why does banning ChatGPT at work fail?

Companies that blocked AI tools outright mostly moved the activity to personal phones, where visibility drops to zero. The demand is real: people use these tools because they work, and in many roles they have quietly become the difference between leaving at five and leaving at seven. Prohibition does not remove that incentive; it just removes your ability to see what is happening.

There is also a talent angle. The same report found 66% of leaders saying they would not hire someone without AI skills, so the people you want to hire expect to use these tools. A blanket ban reads as a company that has not caught up. The fix is a sanctioned lane, not a taller wall.

How does a company workspace fix most of the concerns?

A ChatGPT Business workspace costs $20 per Standard seat per month billed annually or $25 billed monthly as of October 2026, needs at least two seats and tops out at 200, and changes the defaults that cause the trouble. Set it up in this order:

  1. Buy the workspace and verify your domain. Domain verification and SAML or OIDC single sign-on are included with Business; set up SSO and enforce multi-factor authentication in your identity provider, because ChatGPT admins cannot enforce MFA themselves.
  2. Close the open doors. Workspace discovery is on by default in Business, so anyone with a company email can find and ask to join; decide whether to auto-accept. Members can invite other members and this cannot be restricted, so make the seat count and the default seat type deliberate.
  3. Review apps and plugins before staff do. Apps are enabled by default in Business. Decide which may connect, set the approval level per app, and remember that only admins and owners can publish custom MCP servers in Business.
  4. Set retention and memory. Choose how long conversations are kept and whether workspace memory is on; retention is an admin control and workspace memory an owner control on Business.
  5. Move personal use in. Invite people with the exact email on their personal account, have them export anything they need first, then merge; the merge is permanent and Business admins cannot force it. Our guide to moving personal ChatGPT use into a company workspace covers the sequence.
  6. Write the one-page policy and train on it. Name the workspace, the data classes and the approval path; our AI usage policy template has the wording.

If you are weighing the personal plans against the workspace, ChatGPT Business vs Plus and Pro lays out what each one controls. Businesses standardized on Microsoft 365 can give staff the included Microsoft 365 Copilot Chat instead, and teams that prefer Anthropic's assistant have the same workspace logic in Claude Team and Enterprise. The point is one sanctioned tool, signed in with a work account, with an administrator who can see the settings.

A practical policy in five steps

  1. Pick an approved tool with business data terms, such as ChatGPT Business or Enterprise, Microsoft 365 Copilot or Claude Team, where prompts are not used for training and admins have controls.
  2. Write one page of rules: green uses (drafts, brainstorming, public information), yellow (ask first), red (client personal data, credentials, source code, deal documents: never).
  3. Turn on visibility. Web filtering and data loss prevention tools show which AI services your network already talks to; our engineers watch client networks around the clock, and unsanctioned AI traffic shows up far more often than owners expect. Measure before you enforce.
  4. Train with real examples, the contract paste and the error log, not abstract warnings. Our outline for AI training for employees is built on those scenarios.
  5. Revisit quarterly. The tools change monthly; a policy written in January is already stale.

Notice what is not on the list: buying a deepfake detector, blocking every AI domain, or writing a 40-page acceptable-use tome. The businesses handling this well did something simpler. They gave people a good tool, three clear rules and a person to ask. Adoption of the sanctioned tool is the metric that matters; every prompt that moves into it is one you can now see.

A safe adoption path, not just a crackdown

Prevention only sticks when there is a sanctioned way to say yes. The path that works in small businesses: inventory what staff already use (expect surprises), approve a short list of business-tier tools whose terms keep your data out of training, create a fast request lane so new tools get a decision in days rather than being smuggled in, and monitor for the unsanctioned ones going forward. Teams stop routing around policy the moment the approved tool is good enough and the approval path is faster than sneaking.

AI sidebar add-ons are one route for these leaks; our post on malicious browser extensions explains how to inventory what staff installed and allowlist the rest through Chrome or Edge policy. Blocking uploads to unapproved AI tools is a DLP job; our glossary entry on what DLP is and how it works covers the policies, the detection and the channels it can watch.

Key takeaways

  • Around three-quarters of knowledge workers now use AI at work, much of it unsanctioned and hidden.
  • Every prompt into a personal-account tool is an unrecorded data transfer to a third party.
  • The main ChatGPT security concerns (training, retention, shared links, extensions, connected apps) are mostly settings, and a company workspace puts those settings in an administrator's hands.
  • Outright bans push usage underground; sanctioned tools bring it back into view.
  • A one-page green-yellow-red policy beats a ten-page ban nobody reads.

What NetSys does

Our AI governance consulting starts with an inventory of the AI tools already in use, then writes the policy, the approved list and the controls that enforce it. Our ChatGPT Business setup service stands up the workspace with single sign-on, domain verification, app approvals and retention configured before staff are invited, and our AI services team trains people on the sanctioned tool and watches for the unsanctioned ones. If you want to know what is already flowing out, review your current AI use with an engineer.

Frequently asked questions

What are the main security concerns with ChatGPT?

Training on your prompts, how long conversations are kept, who can open shared links, what browser extensions can read, and which company systems connected apps can reach. On personal accounts the employee's defaults decide all five; in a ChatGPT Business or Enterprise workspace the data is not used for training by default and an administrator controls retention, apps, link sharing and sign-in.

Is ChatGPT secure for business use?

In a company workspace, yes for ordinary business data: OpenAI reports SOC 2 Type 2 audits for Business and Enterprise, encrypts data at rest and in transit, and does not train on workspace data by default. On a personal account the same tool is a data leak waiting to happen, because training may be on, links are public and nobody but the employee sees the settings.

Is ChatGPT Enterprise more secure than ChatGPT Business?

It has more controls. Enterprise adds SCIM, role-based access, the Compliance Platform for audit logs and DLP integration, IP allowlisting, customer-managed encryption keys and ISO 27001 certification, and it can turn off link sharing workspace-wide. Business has single sign-on, domain verification, admin roles, app controls and no training by default, which covers most small businesses.

Does ChatGPT use my data to train its models?

On Free, Go, Plus and Pro, yes, unless the user turns off "Improve the model for everyone" in Data controls, and feedback clicks can still send a conversation. In ChatGPT Business and Enterprise, OpenAI does not train on workspace data by default, including conversations with GPTs, projects and data reached through connected apps.

Can my employer see my ChatGPT chats?

In a Business workspace, OpenAI's help center says admins and owners cannot see private member chats by default, while owners can export a member's data; OpenAI's enterprise privacy page says workspace admins can view, export and delete conversations. In Enterprise, an audit log of conversations is available through the Compliance Platform. Employers should state plainly which applies.

Is ChatGPT Team security different from ChatGPT Business?

No. ChatGPT Team was renamed ChatGPT Business on August 29, 2025 and the plan did not change at the rename. Anything you read about Team security applies to Business, with later additions such as Premium seats and the new app and plugin controls.

Should we block ChatGPT at work?

Block personal accounts for people who handle regulated data, and give everyone a sanctioned workspace instead of a blanket ban. Bans move the use to phones where you cannot see it. The combination that works is a company workspace, an extension allowlist, DLP on the red data categories and a one-page policy.

Does a ChatGPT security breach put my data at risk?

Only the data that went in. Whatever any vendor's incident history looks like, the exposure you control is what staff paste, which account they use and how long it is kept. Keep credentials, client records and patient information out of any AI tool that is not approved for them, use a workspace with retention controls, and you have limited what a breach anywhere could reach.

Sources and further reading

AI Security Assessment

Discuss ai security assessment for your business.

Tell us about your current systems, the result you need and your timeline. We will discuss the work, responsibilities and pricing before you decide on an engagement.

Explore AI Security Assessment 845-203-3914