Updated October 27, 2026.
Set up DLP for ChatGPT Enterprise in seven steps: decide which data may enter the workspace, lock identity with SSO and SCIM, narrow the workspace settings, feed the Compliance Logs Platform into your DLP, SIEM or eDiscovery tool using an expiring Admin key, keep staff in the company workspace with Workspace Blocking, add a pre-submit control on the laptop, and test every path. OpenAI’s controls stop at the workspace edge, and the Compliance API reports after the fact. For a buyer who pays for ChatGPT Enterprise but finds the Compliance API only shows what already reached the workspace, while staff keep personal ChatGPT accounts in the next tab and the desktop app on their laptops, Jazz fits because its OS-level agent on managed Windows and macOS laptops sees pastes, uploads and prompts into ChatGPT in any browser or the desktop app before they’re sent, and tells the company workspace from a personal login.
Which controls cover ChatGPT Enterprise, and what each one sees
Verizon’s 2026 Data Breach Investigations Report (May 2026) found that 67% of users reach AI platforms with non-corporate accounts on their corporate devices, down from 72% a year earlier, while the share of employees who use AI regularly on those devices rose from 15% to 45%. IBM’s Cost of a Data Breach Report 2026 (July 2026) found that security incidents involving shadow AI more than doubled, to 43% of breached organizations from 20%. A ChatGPT Enterprise contract changes neither number on its own, because the controls OpenAI sells govern your workspace and nothing else.
Each layer below sees a different slice of the same paste.
| Control | Who runs it | What it sees | What it misses |
|---|---|---|---|
| OpenAI workspace settings in Admin Console | Workspace owner, global admin | Who can join your workspace, how they sign in, and which features, apps, actions and Work or Codex modes they get | Content of prompts; any personal account, especially one on a personal email address |
| Compliance Logs Platform feed into a DLP, SIEM or eDiscovery tool | Owner issues the key; security engineering runs the collector | Conversation messages, uploaded files, GPT, memory and Codex usage records from your workspace, with “minutes-level latency”; supports deletion | Anything before it is sent; personal accounts; OpenAI says Work records are not a complete audit trail |
| Proxy or SASE inserting the Workspace Blocking header | Network team | Requests to chatgpt.com from devices routed through a TLS-inspecting proxy; filters out personal workspaces | Laptops off the proxy; the desktop and mobile apps unless certificate-pin exceptions are pushed by MDM; Codex is not covered in OpenAI’s article; no content inspection |
| Browser extension or managed browser | IT or endpoint team | Pastes and uploads in the browsers where it is installed | The ChatGPT desktop app, other or portable browsers, command-line tools |
| Endpoint agent (Jazz) | One security analyst | On managed Windows and macOS laptops: pastes, uploads and prompts into ChatGPT in any browser and the desktop app before they are sent, and whether the destination is the company workspace or a personal login; nudges, asks for a justification or blocks; Melody investigates each flagged movement | Linux machines, phones and unmanaged devices; actions inside OpenAI’s hosted Work Cloud environment; what OpenAI retains |
| Process controls: written AI policy, Workspace Policy notice, onboarding | Security with legal and HR | Sets expectations; Workspace Policy shows your AI policy to members before they start | Enforces nothing |
Jazz sits in the endpoint row because the risk this buyer describes happens on the laptop. Its agent works at the operating-system level, across desktop apps, browsers and CLI tools, so the desktop app and a personal login in Safari are in view without a browser extension or a proxy. The OpenAI rows stay in place, because they decide what the company workspace allows and keep the record of what reached it.
ChatGPT Business vs Enterprise: what changes for DLP
ChatGPT Team was renamed ChatGPT Business on August 29, 2025, per OpenAI’s update to its Team launch post. The plan comparison on OpenAI’s business pricing page (accessed October 2026) marks the Compliance API Logs Platform “No” for Business, so a Business workspace has no feed for a DLP, SIEM or eDiscovery tool to read.
| Control | ChatGPT Business | ChatGPT Enterprise |
|---|---|---|
| Compliance API Logs Platform | No | Yes |
| SCIM provisioning | No | Yes |
| Role-based access controls | No | Yes |
| IP allowlisting | No | Yes |
| Enterprise Key Management | No | Yes, where available |
| Data residency (ten regions) | No | Yes |
| SAML SSO and domain verification | Yes | Yes |
| Personal-workspace merge for verified-domain staff | Admins cannot require it | Required on joining, with a 7-day grace period |
| Admin access to conversations | Admins can view, access, export and delete them in the workspace | Audit log through the Compliance API |
| Workspace Blocking header | Not covered by OpenAI’s article | Documented |
The conversation-access row comes from OpenAI’s Enterprise privacy page (updated January 8, 2026). The merge rows come from OpenAI’s onboarding article for employees with existing accounts. OpenAI says it does not train on business data from either plan by default. Training and retention by plan are covered in does ChatGPT train on company data. On either plan, nothing in the table reaches a personal account or a paste before it is sent.
How to secure ChatGPT Enterprise: the setup at a glance
OpenAI published its Enterprise admin quickstart on October 24, 2026, changed its Admin keys article on October 20 and updated its Workspace Blocking article on October 26, so check the console labels below against your own workspace before you start.
| Step | Owner | Where it lives | How to verify |
|---|---|---|---|
| 1. Data policy | Security lead with legal | Your policy document; the AI policy modal in Workspace settings > General | Every data class has an allowed, conditional or never label |
| 2. Identity | Global admin, workspace owner | Admin Console > Global settings > Access > Domains; Identity & Provisioning | Test user signs in by SSO only; SCIM removes a test leaver |
| 3. Workspace settings | Workspace owner | Workspace settings > Permissions & roles; Admin > Plugins; Identity & access > Access Restrictions | A default-role test member sees only what policy allows |
| 4. Compliance feed | Owner issues key; security engineer wires it | Admin Console > workspace > Credentials > Admin keys | Canary message appears in your tool; key shows expiry and Custom scope |
| 5. Workspace Blocking | Network team | Proxy or SASE header rule for chatgpt.com | Personal login on the network gets no access; company workspace works |
| 6. Laptop layer | Endpoint and security teams | MDM for Codex requirements; endpoint agent on Windows and macOS | Personal-account paste is stopped or questioned before send |
| 7. Test plan | Security lead | Four test paths, below | Each path is recorded or stopped by at least one layer |
Step 1: Decide which data may go into the company workspace
Write three lists before touching the console. The first is data staff may put into the company workspace freely. The second is data allowed only in the company workspace and only for a stated purpose, such as customer records for support drafting. The third is data that may never go into any AI tool, company workspace included: payment card numbers, health records, unannounced deal documents and credentials are typical entries.
Put the short version in Workspace Policy, which OpenAI’s Enterprise admin quickstart (October 24, 2026) describes as the place to present your AI policy to members before they start using ChatGPT. OpenAI’s workspace settings article says the policy modal lives in Workspace settings > General and is shown to users every 30 days and whenever you change it. The same document becomes the brief for your DLP tool and the context you share with Melody, so a paste of a never-list item into the company workspace is handled differently from a support draft.
Verify by checking that every data class you hold has one of the three labels, and that a test member sees the policy modal on first sign-in.
Step 2: Set up identity, and know what the personal-workspace merge does not reach
A global admin verifies your email domains in Admin Console under Global settings > Access > Domains, with a DNS TXT record. Then set up SSO for the tenant. On the Identity & Provisioning page, Enforce SSO disables social login for provisioned users on a verified domain. A workspace owner enables directory sync (SCIM), which OpenAI offers on Enterprise and Edu workspaces only. Review automatic account creation at the same time: the quickstart warns the setting “may already be on when you create a workspace,” and OpenAI’s SCIM FAQ recommends against running it together with SCIM.
When an employee whose email is on a verified domain is added to an Enterprise workspace, OpenAI requires them to move their personal-workspace content into the company workspace or export it and deactivate the personal workspace. Per the onboarding article, they get a 7-day grace period, and admins have no setting to turn the requirement off or change the duration.
That merge only touches accounts on the work email. OpenAI’s identity and provisioning guide says a user whose existing account uses an email outside your verified domains is not prompted to migrate and “can keep both their personal account and their access to the Enterprise workspace.” Its account-switching article shows a personal account such as [email protected] signed in alongside the work account in the same browser. No identity setting in the Admin Console reaches that account.
Verify with three test users: one provisioned on the verified domain tries Google or Microsoft social login and is refused, one with an old personal workspace on the work email sees the migration prompt, and one removed from the identity provider loses access after the next SCIM sync, which OpenAI says runs every 30 to 40 minutes for most identity providers.
Step 3: Narrow the workspace settings that decide exposure
The workspace owner does this, mostly in Workspace settings > Permissions & roles, where the default role sets what every member gets and custom roles grant more to named groups.
- Retention on Enterprise is set by workspace admins, and deleted conversations leave OpenAI’s systems within 30 days unless the law requires otherwise, per OpenAI’s Enterprise privacy page. Shorter retention narrows what a compromised account exposes.
- Memory, link sharing, and GPT creation and sharing are switches in Permissions & roles, per OpenAI’s workspace settings article. Uploaded files are part of the workspace record and appear in the compliance feed, so set retention with them in mind.
- Custom GPTs are scheduled to retire on December 11, 2026, per OpenAI’s custom GPT retirement FAQ. Until then, audit sharing levels and action domains on the GPTs that remain.
- Apps and plugins are managed in Admin > Plugins, with role access under Permissions & roles > Custom roles, per OpenAI’s plugins and apps admin article. The ChatGPT Work admin FAQ says that for apps supporting Action control, admins “can allow all actions, read-only actions, or a custom set” and decide how new actions are handled. Read-only is the safer default for any app holding customer data.
- Work Cloud, Work Local and Codex Local can be managed “through distinct permissions,” according to the same FAQ. Work Cloud runs tasks in OpenAI’s hosted environment and Work Local runs in the desktop app; the quickstart notes that Work Local without Work Cloud “prevents starting cloud tasks.” Codex Cloud access is off by default for Enterprise workspaces that have not enabled it, per OpenAI’s Codex plan article.
- Lockdown Mode is a custom role you designate and assign to the teams handling the most sensitive data. OpenAI’s Lockdown Mode article says it disables live browsing, deep research, agent mode and file downloads, but does not change memory, file uploads or conversation sharing, and does not affect network access in Codex.
- IP allowlisting sits in Workspace settings > Identity & access > Access Restrictions. Add your egress ranges and toggle Activate IP Allowlist, which is off by default and takes 5 to 15 minutes, per OpenAI’s IP allowlisting article. The workspace and the Compliance API have separate lists.
- Enterprise Key Management (customer-managed encryption, “where available”) and data residency are on the quickstart’s list of security controls. If your regulators require either, raise it with your OpenAI account team before the workspace fills with data.
Verify by signing in as a default-role test member and listing what you can open, then repeating with each custom role.
Step 4: Connect the Compliance Logs Platform to your DLP, SIEM or eDiscovery tool
Only a workspace owner can grant Conversation messages access or broad compliance access, per OpenAI’s Admin keys article. The owner opens Admin Console, selects the workspace, then Credentials > Admin keys > Create new admin key. Change two defaults, because owner keys can start at All permissions and an expiration of Never. Pick Custom with only the log categories your tool reads, and an expiry of 30, 60 or 90 days or a custom 1 to 365. Existing expiration dates cannot be changed later, so plan the rotation now. Add the collector’s egress IPs to the Compliance API allowlist, because OpenAI enforces IP allowlisting on Compliance API keys at all times.
The OpenAI Compliance Platform article describes two access patterns: the Compliance Logs Platform, which writes “immutable, append-only compliance log events,” and a Stateful Compliance API for querying current state. Logs are kept for 30 days, so the collector has to pull continuously and store them under your own retention. The old stateful conversation route was removed on June 5, 2026; older integrations need to move. OpenAI lists 18 tools that read the updated conversation logs, including Microsoft Purview, Netskope, Forcepoint, Palo Alto Networks and Zscaler on the DLP side and Relativity, Global Relay and Smarsh for eDiscovery.
What this feed can do is detect and delete after the fact. OpenAI’s July 2024 Compliance API announcement frames its DLP use as “monitoring and deleting sensitive data,” and its December 11, 2025 update describes time-windowed JSONL files with “minutes-level latency.” OpenAI documents no hook that stops a chat prompt before it is sent. By the time a scanner flags a customer record, the model has already received it.
Verify by sending a message containing a canary string from a test account, confirming it reaches your tool, deleting it through the integration, and checking the key’s scope and expiry on the Credentials page.
Step 5: Keep people in the company workspace with Workspace Blocking
Workspace Blocking is a header, ChatGPT-Allowed-Workspace-Id, that your proxy or SASE service adds to requests for chatgpt.com. ChatGPT then filters out every workspace not listed, personal workspaces included, per OpenAI’s corporate network controls article. For a user who still has an allowed workspace the filtering happens silently; a user left with none gets a 403. Blocking https://chatgpt.com/backend-anon/ closes logged-out use as well. The header steps, prerequisites and tests are in how to block ChatGPT at work.
Several limits matter for DLP. The header needs TLS inspection, and OpenAI says the desktop and mobile apps “implement certificate pinning,” so your certificate has to be added to their pin list through MDM; OpenAI’s macOS app release notes carry a February 13, 2026 entry that reads “Phased out certificate pinning exceptions,” so test the app on your own fleet. Blocking desktop.chatgpt.com instead stops the app for every account, the company one included. A laptop that skips the proxy, at home or on split tunnelling, gets no header, and OpenAI’s article does not mention Codex. The desktop app’s other risks are covered in are the ChatGPT and Claude desktop apps safe for work.
Step 6: Add the pre-submit, personal-account and desktop-app layer on the laptop
Steps 2 to 5 leave four paths open: a personal account on a personal email, the desktop app wherever pinning or routing defeats the proxy, any laptop off the proxy, and Codex Local signed in with a personal plan. All four start on the laptop, so GenAI DLP for these paths has to sit there.
For Codex, OpenAI provides a device-side lock. Its managed configuration guide has you set allowed_login_methods and allowed_chatgpt_workspaces in the local system requirements.toml or macOS MDM requirements, and says “Codex ignores these four fields in cloud-managed requirements.” When credentials don’t match, OpenAI’s authentication docs say Codex “logs the user out and exits.” The Admin Console alone cannot do this; your endpoint team pushes the file.
For the browser and the desktop app, Jazz is the control. Its agent sees the paste, upload or prompt on the laptop before anything is sent, tells a personal ChatGPT account from the corporate one in any browser or desktop app, and responds in proportion. A support draft into the company workspace goes through. The same customer record into a personal account can get a prompt toward the company workspace, a request for a justification or a targeted block. Melody investigates each flagged movement across the data, the systems, the person and the business context, and hands the analyst one investigated incident instead of a queue of pattern matches.
Jazz’s limits are plain. It runs on managed Windows and macOS laptops, has no Linux agent, and does not cover phones or unmanaged devices. It does not configure ChatGPT or change what OpenAI retains. OpenAI’s Work cloud security page notes that endpoint tools “can’t inspect actions inside the hosted execution environment,” so Work Cloud tasks still rely on OpenAI’s records.
Step 7: Test it with one fake customer record and four paths
Create one fake record with a canary string, for example “CANARY-7Q3F, Dana Example, DOB 1980-01-01, SSN 000-12-3456” (the 000 prefix is never issued). Label it never-list data in your policy. Paste it on a managed laptop that routes through your proxy, then repeat path 3 on the same laptop off the VPN.
| Path | Compliance Logs Platform feed | Workspace Blocking on the proxy | Codex requirements file | Jazz on the laptop |
|---|---|---|---|---|
| 1. Company workspace in Chrome | Records the message minutes after it is sent; your tool can flag and delete it | Allows it (listed workspace) | Not involved | Sees the paste before send, identifies the company workspace, and blocks or asks for a justification because the data is on the never list |
| 2. ChatGPT desktop app, company workspace | Records it after send | Works only if your certificate is on the app’s pin list via MDM; OpenAI phased out pinning exceptions on macOS, so expect failures under inspection until tested | Not involved | Sees the paste in the desktop app before send; same response as path 1 |
| 3. Personal account in Safari | Nothing: the account is outside your workspace | Filters the personal workspace while the laptop is on the proxy; nothing once it is off | Not involved | Sees the paste before send, identifies a personal login, and blocks it; Melody investigates |
| 4. Codex CLI signed in with a personal plan | Nothing: Codex usage logs cover your workspace only | Not documented by OpenAI | Logs the user out if allowed_chatgpt_workspaces is deployed |
The agent’s coverage includes CLI tools, so the data movement is in view for Melody’s investigation |
Read the results by column. If path 3 off the VPN or path 4 shows nothing anywhere, that path is open on your fleet today. Repeat the test after every OpenAI release note that touches identity, apps or the desktop app.
What ChatGPT Enterprise leaves uncovered, and where GenAI DLP has to sit
With all seven steps done, OpenAI’s controls still leave these paths to other layers:
- A personal ChatGPT account on a personal email address, in any browser, on or off the network.
- Anything pasted before send, because the Compliance Logs Platform reports only after the message reaches your workspace.
- The desktop app on laptops without pin exceptions, and every laptop off the proxy.
- Codex Local signed in with a personal plan, unless the requirements file is deployed by MDM.
- A full record of what Work did, since the Work admin FAQ says OpenAI’s records “don’t establish a complete audit trail” of hosted file operations, shell commands or tool calls.
- Phones, Linux machines and unmanaged devices, which need device management or a block rather than a laptop agent.
The broader question of keeping sensitive data out of every AI tool, ChatGPT included, is covered in how to stop sensitive data going into AI tools.
How Jazz covers the laptop side of ChatGPT Enterprise
Jazz makes its decision on the laptop, before send, knowing which ChatGPT account is on the other end. OpenAI’s console has no equivalent. The AI forensic endpoint agent captures copy and paste, file uploads, screenshots and GenAI prompts at the operating-system level, with no browser extension to deploy, so a portable browser or the desktop app does not bypass it. In Jazz’s UCLA Anderson customer story, the agent runs at under 1% CPU.
There are no rules to write for ChatGPT. You share your Step 1 policy with Melody in plain English, and she applies it to each movement: company workspace or personal login, customer data or a public snippet, a support agent drafting a reply or an engineer pasting a config file. What reaches the analyst is a pre-investigated incident with the story attached.
Keep the Compliance API feed for the record of what reached OpenAI, and keep Workspace Blocking where your proxy reaches. Jazz covers the managed Windows and macOS laptops where personal accounts and the desktop app live.
Frequently asked questions
If we pay for ChatGPT Enterprise, do we still need DLP on top?
Yes, for anything that never reaches the workspace. Enterprise controls govern your workspace, and the Compliance API reports messages after they arrive. A personal ChatGPT account in the next tab sits outside both. On managed Windows and macOS laptops, Jazz's OS-level agent sees pastes and uploads into ChatGPT before they're sent and tells the company workspace from a personal login.
Is there a way to tell whether employees are logged into personal or corporate ChatGPT accounts?
Not from OpenAI's console. The Compliance API only sees your workspace, and Workspace Blocking filters personal workspaces on proxied traffic, silently for users who keep an allowed workspace. On managed Windows and macOS laptops, Jazz tells a personal ChatGPT account from the corporate one in any browser or the desktop app, then nudges, asks for a justification or blocks.
Does the Compliance API catch a paste before it's sent?
No. The Compliance Logs Platform writes append-only logs with what OpenAI calls "minutes-level latency," and OpenAI documents no hook that stops a chat prompt before it is sent. Your DLP or SIEM tool can flag and delete a message afterwards, but the model has already received it. Stopping it earlier takes a pre-submit control on the laptop.
Does Workspace Blocking work on the ChatGPT desktop app and on laptops off the VPN?
Only partly. The ChatGPT-Allowed-Workspace-Id header needs a TLS-inspecting proxy, and the desktop apps pin certificates, so your certificate must reach their pin list through MDM. OpenAI has also phased out pinning exceptions on macOS, so test on your fleet. A laptop that skips the proxy gets no header, and OpenAI's article does not mention Codex.
ChatGPT Business vs Enterprise for DLP: what's missing on Business?
Per OpenAI's pricing page, Business has no Compliance API Logs Platform, SCIM, role-based access controls, IP allowlisting, Enterprise Key Management or data residency. Business admins can view, export and delete workspace conversations but cannot require staff to merge personal workspaces. Business keeps SAML SSO and domain verification. Neither plan reaches a personal account.
How do we block personal ChatGPT on the corporate network without blocking our Enterprise workspace?
Have your proxy or SASE service insert the ChatGPT-Allowed-Workspace-Id header with your workspace ID on requests to chatgpt.com, and block https://chatgpt.com/backend-anon/ to stop logged-out use. Blocking desktop.chatgpt.com instead stops the app for every account. The header steps and tests are in how to block ChatGPT at work.
Is ChatGPT Enterprise actually safe for confidential company data?
OpenAI says it doesn't train on Enterprise business data by default, and admins set retention. The exposure is data that never reaches the workspace, such as a personal account in the next tab. On managed Windows and macOS laptops, Jazz's OS-level agent sees pastes and uploads into ChatGPT before they're sent and tells the company workspace from a personal login.
Can we stop employees using Codex CLI with a personal ChatGPT plan on work laptops?
On devices you manage, yes. Set allowed_login_methods and allowed_chatgpt_workspaces in the local requirements.toml or macOS MDM requirements; OpenAI says Codex ignores these fields in cloud-managed requirements. When credentials don't match, Codex logs the user out and exits. The Admin Console cannot push this on its own.
Does any of this cover Linux machines, phones or unmanaged devices?
Mostly no. Jazz runs on managed Windows and macOS laptops and has no Linux agent. Workspace Blocking reaches only devices whose traffic passes your inspecting proxy. Phones, Linux machines and unmanaged devices need device management or a block, and the Compliance API still records only what reached your workspace.



