Is Grok Safe for Confidential Information? Grok on X, grok.com, Business, Enterprise and the API

Is Grok safe for confidential information? Not on the consumer service or inside X: conversations can be used to train models unless you opt out, authorised staff may review them, and the company's own FAQ tells users not to share personal or sensitive information. Grok Business, Enterprise and the API state no training and 30-day deletion, with zero data retention available on the API. All of them are cloud services.

Four routes lead to the same models. Grok inside X, the social network, sits under X’s terms and privacy policy. The grok.com website and the Grok mobile apps, free or with a SuperGrok subscription, sit under the developer’s consumer terms. Business and Enterprise plans and the API sit under its Enterprise terms. One naming point: the legal pages at x.ai now call the company SpaceX Super Intelligence LLC, shortened to SpaceXSI, and say it remains a separate company from X Corp. This guide calls it xAI, the name most people still search for.

We read the x.ai legal pages, the API security documentation, the Business plan page and X’s privacy policy on 14 November 2026, and checked the regulator material on the Irish Data Protection Commission’s website. One page could not be read directly: X’s help centre article on Grok sits behind a bot check, so the menu path we give for X is the one the article is reported to show and should be confirmed in your own account.

Grok plans and data handling at a glance

Is Grok safe to use on grok.com and in the apps?

For questions that contain nothing sensitive, yes. The consumer FAQ is direct about the rest. Under the heading on model training it says the company may use your content and interactions, meaning prompts, searches and other material you submit, along with Grok’s responses, to train its models. In the same answer it asks users to protect their privacy by not sharing personal or sensitive information in their questions.

Three more points from the same pages matter for anyone thinking of pasting work material:

• People can read conversations. A limited number of authorised personnel may review conversations for specific business purposes, which the FAQ lists as improving model performance, investigating security incidents and potential misuse, and meeting legal obligations.
• Deletion takes up to 30 days and has exceptions. Deleted conversations are removed from the company’s systems within 30 days unless they have been de-identified or pseudonymised and separated from your account, or must be kept for safety, security or legal reasons.
• Related companies can receive data. The privacy policy, last updated on 27 October 2026, lists contracted service providers, business transfers, legal purposes and related companies among the recipients of most categories of personal information.

The policy covers the Grok apps and grok.com only. It says outright that it does not apply to data processed for business customers, and that use of Grok on the X platform is governed by X’s privacy policy and terms instead. A paid SuperGrok subscription raises limits; nothing on the pages we read gives it different data terms from the free tier.

Does Grok train on your data? The settings that control it

On the consumer service you choose, and the default is yes. The FAQ gives the paths:

1. Mobile app: Settings, then Data Controls, then switch off “Improve the model”.
2. grok.com: Settings, then Data, then “Improve the Model”.
3. Per conversation: start a Private Chat instead.

The FAQ adds the limits. Opting out applies to new conversations. Feedback you choose to send, such as rating a response, may still be used for training. If you use Grok without logging in, the company may collect content on an anonymous basis, and in some regions outside the EU and UK there is then no opt-out at all. So sign in before you rely on the setting.

A second switch, “Personalize Grok using X”, controls whether your X data and interactions shape Grok’s answers. It is separate from training and worth checking on its own.

What Private Chat does, and what shared links expose

Private Chat is the ghost icon at the top right of the screen. The company says conversations in this mode do not appear in your history, are not used for model training, and are deleted from its systems within 30 days unless they have been de-identified or must be kept for safety, security or legal reasons. That is narrower than the name suggests: the text still travels to the service and is stored for up to a month.

Shared links are the opposite of private. The share button creates a public link, and the FAQ warns that a link posted publicly may be indexed by a search engine like any other public content. This is not hypothetical. In August 2025, Forbes reported that a Google search showed more than 370,000 Grok conversations indexed, and that the share URLs were also available to Bing and DuckDuckGo. The published conversations included personal details, at least one password, and uploaded spreadsheets, images and text documents. According to Forbes, users were given no warning that sharing made a chat searchable, and the company did not respond to its request for comment.

You can see and revoke your own links at grok.com/share-links. For work, the simpler rule is not to share a conversation that contains anything you would not publish.

Is Grok in X safe? Posts, interactions and the X opt-out

Grok inside X is the least suitable route for confidential material, because two sets of data are in play: what you type to Grok, and what you post on X.

X’s privacy policy says X may use the information it collects, and publicly available information, to help train its machine learning or artificial intelligence models. Its section on third-party collaborators says that, depending on your settings, X may share your information with third parties, and that if you do not opt out the recipients may use it for their own purposes, including training their AI models.

The opt-out is an X setting, not a Grok one. X’s help article is reported to place it under Settings, Privacy and safety, Data sharing and personalization, Grok and Third-party Collaborators. X changes its menus often, so look for the Grok entry under the privacy settings of your own account.

Regulators have looked closely at this route:

• August and September 2024. Ireland’s Data Protection Commission (DPC) went to the High Court over X’s use of EU and EEA users’ public posts to train Grok. The Law Society Gazette reported on 4 September 2024 that the proceedings were struck out after X agreed to adhere permanently to its earlier undertaking to suspend that processing.
• 11 April 2025. The DPC announced an inquiry into X Internet Unlimited Company, examining whether personal data in publicly accessible posts by EU and EEA users was lawfully processed to train the Grok models.
• 17 February 2026. The DPC opened a second, large-scale inquiry, this time into the creation and publication on X of non-consensual intimate or sexualised images using Grok’s generative functions. It covers GDPR Articles 5, 6, 25 and 35. We found no published decision in either inquiry when we checked.

None of these is a finding that Grok leaks business data. They do show that the legal basis for training on X data is contested in the EU. Our guide to AI tools and the GDPR explains what a controller has to document before adopting any hosted assistant.

Is Grok safe for confidential information on Business or Enterprise?

These plans exist today on the company’s business page, under the names Business and Enterprise, and they change the terms rather than only the limits. Both list “No training”. The Enterprise FAQ repeats it: the company does not use business data, inputs or outputs, to train its models, though it may offer free credits in exchange for permission to do so. Decline that offer if it is made for an account that handles confidential work.

What each plan lists:

• Business: SOC 2 Type I and II, role-based access control, team and seat management, domain verification and claims, user analytics, custom data retention, connectors, and advanced audit and security controls.
• Enterprise adds: custom single sign-on, SCIM directory sync, custom roles, customer-managed encryption keys, application-level encryption and a dedicated data plane.

Single sign-on and directory sync are Enterprise features, so a Business team manages accounts by hand. The page also describes an “Enterprise Vault” in which data is encrypted with the customer’s keys and isolated from other customers, and lets teams attach connectors to tools such as Slack, Salesforce and Notion or bring their own MCP server. Each connector widens what Grok can read, so approve them one at a time.

On contracts, the Enterprise FAQ says the data processing addendum is incorporated automatically into the Enterprise terms when a customer submits personal data. On health data, its wording is cautious: “under some circumstances” the company can support a customer’s HIPAA obligations, and a business associate agreement starts with a questionnaire. The business page shows SOC 2 Type II, HIPAA, and GDPR and CCPA badges; treat those as the vendor’s statements and ask for the reports. The API documentation says the Trust Center is available to customers with a signed non-disclosure agreement.

Two cautions. The Enterprise FAQ was last updated on 25 February 2025, so confirm the current terms in the contract. And staff review is not ruled out: a limited number of authorised personnel may review business data when legally required, for example to investigate security incidents or misuse.

Is the Grok API safe? Retention, zero data retention and regions

The API has the clearest terms of the four routes. The security FAQ in the developer documentation says the company never trains on API inputs or outputs without explicit permission. Requests and responses are stored for 30 days, encrypted at rest, so suspected abuse can be audited, and are then deleted automatically.

Zero data retention (ZDR) removes that 30-day store. Where self-serve ZDR is available, a team admin enables it in the Console under Team Settings. It applies to the whole team, not to a single key. Existing Files and Collections must be deleted first, and features that need stored data stop working: stateful Responses API calls, file uploads, Collections, the Batch API, deferred completions and stored image or video outputs. The documentation says most customers should not enable it unless compliance requires it, because content deleted under ZDR can never be recovered. Every API response carries an x-zero-data-retention header set to “true” or “false”. To see it, send the request from the documentation’s quickstart and print the response headers instead of the body:

curl -s -D - -o /dev/null https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "grok-4.7", "input": "ping"}' \
  | grep -i x-zero-data-retention

Location depends on the endpoint. The default, https://api.x.ai, does not guarantee a processing region. A US endpoint, https://us.api.x.ai/v1, keeps request handling, inference, moderation and retained request data in the United States. It serves only the grok-4.7 and grok-4.6 models, has no image, video or voice APIs, and costs 10% more per token; files, Collections and server-side tools fall outside the guarantee. The documentation we read describes no EU endpoint, so an organisation that needs inference inside the EU cannot get it from this API today. Our zero data retention explainer sets these terms beside other providers’.

Is Grok safer than ChatGPT?

For a consumer, the written terms are similar. Both services can use chats to improve their models by default, both let each user opt out, and both have a temporary mode that keeps a conversation out of history and training. Our plan-by-plan ChatGPT guide has OpenAI’s side.

Three differences are worth weighing:

• The social network. Grok is built into X, where public posts and Grok interactions come under a second policy with its own opt-out. ChatGPT has no equivalent source of user posts.
• Regulatory history. The DPC proceedings and inquiries listed above concern Grok’s training data and outputs directly.
• Business tooling. Single sign-on and SCIM sit only on Grok’s Enterprise plan, and the API offers a US-only endpoint but no EU one. Compare that with the specific ChatGPT plan you would buy.

The honest answer for an employer is that neither consumer service is suitable for confidential data, and that the business plans should be compared contract by contract. The same exercise for other assistants is in our guides to Claude’s consumer and commercial plans, Copilot on personal and work accounts and Gemini across Google’s products.

What to do about Grok: employees and IT

If you use Grok yourself:

1. Keep client data, credentials, source code and personal data about other people out of grok.com, the apps and Grok in X.
2. Sign in, then switch off “Improve the model”. Check the Grok entry in X’s privacy settings as well.
3. Use Private Chat for anything you would not want in your history, knowing it is still stored for up to 30 days.
4. Do not use the share button on a conversation that includes a document or a name. Review grok.com/share-links and remove old links.
5. Do not send feedback on a response that contains sensitive text.

If you run IT or security:

1. State in your AI acceptable use policy which Grok route, if any, is approved for which class of data.
2. If Grok is approved, buy Business or Enterprise, refuse any credits-for-training offer and get the data processing addendum on file.
3. For API use, decide on ZDR per team, use the US endpoint where residency is required, and log the response header as evidence.
4. Approve connectors and MCP servers individually, and prefer read-only scopes.
5. Treat personal accounts on managed devices as unapproved use and monitor for them; our shadow AI explainer covers detection.

When Grok is the wrong place for the data

Every route above ends in infrastructure the vendor operates. A dedicated data plane and customer-managed keys narrow who can read the data; they do not move the processing into your building. Some obligations require exactly that: client contracts that forbid third-party processing, export-controlled or classified work, and EU organisations that need inference inside the EU. Grok’s weights are not offered for self-hosting on the pages we read, so running it yourself is not an option.

The alternative is an open-weight model on hardware you control, tested on your own tasks before anyone depends on it. Our comparison of private and public LLMs sets out what you gain and what you take on.

How VDF AI fits

VDF AI does not run, host, resell or integrate with Grok, on X or anywhere else. It is a separate, self-hosted platform for chat, retrieval over company documents and governed agents that an organisation deploys on-premises, in a private cloud or air-gapped, keeping conversation history and audit records in its own storage.

The difference from one person changing a privacy setting is who decides. In VDF AI Chat an administrator assigns a model to each agent, and VDF AI Router registers on-premises models and can restrict routing to them in air-gap mode. Role-based access control applies on every plan, so access follows the organisation’s rules instead of each user’s settings. To see whether that suits your data, get in touch.