How to Block Personal ChatGPT and Claude Accounts at Work
Keep corporate ChatGPT and Claude available while restricting personal accounts. Compare tenant restrictions, managed Chrome controls, and DLP.

Rushali
2026 State of Agentic AI Security
How ready is your organization to secure AI agents?
3 quick questions. See how you compare to other security leaders.
You can block personal ChatGPT and Claude accounts at work without blocking ChatGPT or Claude. IT allows the approved company workspaces, restricts personal-account access on the traffic it controls, and applies DLP inside the approved sessions that remain. A domain block cannot do this, because chatgpt.com and claude.ai serve personal and Enterprise accounts from the same hostnames.
The distinction that matters is which account versus what data. Tenant controls decide which account a session belongs to. DLP decides what is allowed into it. You need both, and each one stops at a boundary worth knowing before you roll it out.

Control | Purpose | Coverage limit |
|---|---|---|
Provider tenant restrictions (OpenAI Workspace Blocking, Claude Tenant Restrictions) | Allow only your approved ChatGPT workspace or Claude organization; personal accounts are filtered or refused | Works only on traffic your proxy or SASE decrypts and injects a header into; unmanaged networks and devices bypass it |
Akto Atlas managed Chrome account controls | Block personal accounts on chatgpt.com and claude.ai inside managed browsers, on or off the corporate network | Covers browsers where the Akto extension is deployed; not desktop apps, mobile apps, or unmanaged browsers |
DLP in approved sessions | Block or redact PII, secrets, and source code in prompts sent to approved workspaces | Covers supported sessions where inspection runs; it does not decide which account is in use |
Jump to your provider:
Block Personal ChatGPT, Allow ChatGPT Enterprise: Workspace Blocking Filters by Workspace, Not Domain

OpenAI's control is called Workspace Blocking, and it is a ChatGPT Enterprise feature. Your proxy or SASE service adds a custom HTTP header, ChatGPT-Allowed-Workspace-Id, to requests bound for https://chatgpt.com/*. The header carries the UUID of each workspace you approve. ChatGPT then hides every workspace not on the list, personal workspaces included.
The filtering is quiet when it works. If a user still has at least one allowed workspace after filtering, the personal one simply disappears. If filtering leaves them with none, they get a 403 Forbidden. A malformed header (a typo in the UUID, a space after a comma) fails every request with a 400 Bad Request, so test the header value before you roll it out wide.
Where the Workspace ID comes from
Sign in as a ChatGPT Enterprise admin, open Admin Settings, and copy the Workspace ID (a UUID) for each workspace you want to allow. Multiple workspaces go in the same header, separated by commas with no spaces.
Where IT configures it
The header is injected wherever you already inspect TLS: a secure web gateway, a forward proxy, or a SASE platform. OpenAI points enterprise-wide deployments to SASE partners, and also allows a managed browser configuration that inserts the header. No inspection, no header, no control.
Logged-out ChatGPT is a separate hole
ChatGPT works without any account at all. Workspace Blocking cannot filter a session that has no workspace, so block URLs that begin with https://chatgpt.com/backend-anon/ in the same gateway or browser policy. Skip this and a user who signs out has an unfiltered path.
Desktop and mobile apps need more than the header
The ChatGPT desktop and mobile apps use certificate pinning, which breaks under TLS inspection. To enforce Workspace Blocking in those apps, add your inspection certificates to the pin list and push them through MDM. That only works on managed devices.
The blunt alternative is blocking the app domains outright (android.chat.openai.com, ios.chat.openai.com, desktop.chatgpt.com). OpenAI is clear that this blocks everyone on those apps, Enterprise users too.
Key takeaway: Inject the workspace header on chatgpt.com, block the anonymous path, and treat desktop and mobile as a separate MDM project. Follow OpenAI's Corporate Network Controls in ChatGPT Enterprise guide for the exact setup.
Block Personal Claude, Allow Claude Enterprise: Tenant Restrictions Hold Only on Your Network Path
Anthropic's equivalent is Claude Tenant Restrictions, available to Enterprise plan and Console organizations. Your proxy injects an anthropic-allowed-org-ids header into Claude traffic. Anthropic checks the authenticated user's organization against that list and refuses anything outside it with a 403 tenant_restriction_violation error.
Coverage is broader than the browser tab. Anthropic lists web access on claude.ai, desktop and app access, API key authentication, and OAuth token authentication as supported. The restriction also applies to artifacts, including Claude Docs.
The approved organization ID
Enterprise members find the Organization ID under Settings > Account, or at the bottom of Organization settings > Organization. Console organizations find it under Settings > Organization. Several approved orgs (a partner tenant, a second business unit) go in one comma-separated header with no spaces, up to 500 UUIDs across at most 10 header lines.
Where the restriction applies
Scope the proxy rule to claude.ai, api.anthropic.com, claude.com, and anthropic.com, and turn TLS inspection on; Anthropic marks it required. Configure the proxy to overwrite the header on every request rather than append one. A duplicated header returns a 400, and so does a malformed continuation header.
Anthropic documents tested setups for Cato Networks, Cloudflare Zero Trust, Netskope, Palo Alto Prisma Access, and Zscaler ZIA, plus any HTTPS proxy that can insert headers.
Verify the Claude clients you actually ship
The same help article is explicit about the boundary: standard authentication continues on unmanaged networks. A laptop on home Wi-Fi without your agent or tunnel never sees the header. Test each client you deploy (browser, Claude Desktop, mobile, Claude Code) from inside the restricted path, and confirm that each one actually routes through it.
Key takeaway: Tenant Restrictions are strong on traffic you inspect and silent everywhere else. Follow Anthropic's Enforce network-level access control with Tenant Restrictions guide, then map which devices never hit your proxy.
Block Personal AI Accounts in Managed Chrome: Akto Atlas Enforces Inside the Browser
Both provider controls depend on a proxy seeing decrypted traffic. Akto Atlas takes the opposite route for browser use: the Akto Chrome extension evaluates requests inside managed browsers, against policies you define centrally, before the request completes. That makes it useful on managed laptops that leave the office network, and on devices where you have not finished TLS inspection rollout.
The Block personal accounts guardrail
Atlas ships a Block personal accounts guardrail. Akto's documentation lists chatgpt.com and claude.ai as supported, alongside gemini.google.com, copilot.microsoft.com, and grok.com. It detects sign-ins that use personal email domains and restricts access to organization-approved accounts. You can also scope other guardrails to personal and unidentified accounts while excluding work accounts.
Enforcement has two modes, set by the Bypass Guardrails setting. With bypass on, violations are logged and the request proceeds; use this to pilot. With bypass off, the request is blocked, the user sees a browser alert, and the site returns a request failure. End users cannot change this setting. Configure scope and modes using the Akto Chrome extension documentation and the extension usage behaviour guide.
Keep the claim scoped
This covers the web apps in browsers where the extension is deployed: Chrome and Chromium-based browsers, with separate Firefox and Safari extensions. It does not cover the ChatGPT or Claude desktop apps, which Akto handles through a different Atlas component (AI Endpoint Shield), and it does not cover an unmanaged browser on a personal device.
DLP for what employees type into approved sessions
Blocking the personal account solves the wrong-tenant problem. It does nothing about an employee pasting a customer list into the approved Enterprise workspace. Atlas guardrails inspect prompts in supported sessions and block or redact PII, secrets, source code, and other regulated data before they reach the model. Every flagged event, blocked or allowed, lands in Akto Atlas > Agentic Guardrails > Guardrail Activity with the user, path, and triggering policy. See the AI DLP FAQ for how the DLP policies work.
Key takeaway: Use Atlas to enforce account and data policy on managed browsers, including off-network. Book an Akto Atlas demo to see the Block personal accounts guardrail on your own ChatGPT and Claude tenants.
Rollout Checklist: Four Tests Before You Call It Done

Run every test from a managed device on the enforced path, with one approved test user and one personal test account per provider. A control that has not failed a personal login in front of you has not been tested.
Check | How to test | Expected result |
|---|---|---|
1. Company session works | Sign in to chatgpt.com and claude.ai with the approved work account | The approved workspace or organization loads normally; no 400 errors from a malformed header |
2. Personal session is blocked on the configured surface | Sign in with a personal account in the same browser or app | ChatGPT hides the personal workspace or returns 403; Claude returns |
3. Synthetic sensitive data triggers DLP | Paste a fake secret or a test card number (such as 4111 1111 1111 1111) into the approved workspace | The prompt is blocked or redacted according to your DLP rule |
4. The guardrail event appears | Open Akto Atlas > Agentic Guardrails > Guardrail Activity | Events for tests 2 and 3 show the user, endpoint, path, and triggering policy |
Validate desktop apps, mobile apps, and unmanaged paths separately wherever they are in scope; a pass in Chrome says nothing about them.
Copyable Employee Policy and FAQ
Technical controls work better when employees know the rule they enforce. Paste this into your acceptable-use policy or onboarding page:
Use company-managed ChatGPT or Claude accounts for work. Do not enter company data into personal AI accounts.
That line is the minimum. For what else a full policy should cover, see what belongs in an enterprise AI acceptable-use policy.
Can a domain block tell personal accounts from Enterprise accounts?
No. Personal and Enterprise users reach the same hostnames, chatgpt.com and claude.ai. Blocking those domains blocks everyone. OpenAI states the same for its app domains: blocking them stops Enterprise users along with personal ones. Telling accounts apart needs a header-based tenant control or an in-browser account check.
Does signing in with a work email prove the employee is in the approved workspace?
No. OpenAI notes that one account can hold separate personal and company workspaces, so a work email can still open a personal workspace. That is why both providers check a workspace or organization ID, not the email domain.
Does an approved workspace still need DLP?
Yes. Tenant controls guarantee the data lands in your workspace, under your contract and retention settings. They do not stop an employee from pasting credentials, customer records, or source code into it. DLP covers what goes in.
Do these controls cover the ChatGPT and Claude mobile apps?
Only with extra work. ChatGPT's apps need certificate-pinning exceptions pushed through MDM before Workspace Blocking applies. Claude lists desktop and app access as supported, but only traffic routed through your proxy carries the header. Test each app you allow.
The Common Thread: Control Follows the Path You Own
Every control here enforces on a path you manage: the proxy, the managed browser, the inspected session. To block personal ChatGPT and Claude accounts reliably, map the paths first, then cover each one.
Discovery comes first, because you cannot block an account path you have not found. Our workforce AI governance guide covers how to inventory employee AI use before you enforce.
Inventory where employees reach ChatGPT and Claude: browsers, desktop apps, mobile, and Claude Code.
Collect your ChatGPT Workspace ID and Claude Organization ID.
Inject both headers at your proxy or SASE, and block the ChatGPT anonymous path.
Deploy the Atlas Chrome extension in monitoring mode, then switch to blocking.
Turn on DLP for approved sessions and run the four-test checklist.
Publish the employee policy.
The personal account was never the real risk. The unmanaged path to it is.
Experience enterprise-grade Agentic Security solution
