A sales rep asks Grok to turn a long customer thread into a clean summary: names, the account, the deal, the complaint. Pleased with the result, he taps "share" to send it to a colleague. The link works perfectly.
What he does not realise is that, in 2025, that share link was not a private transfer. It was a public web page. Hundreds of thousands of Grok conversations shared this way were indexed by Google, turning what people thought were private chats into searchable documents, complete with names, personal details and at least one password. There was no warning. The button just said "share".
That is the thing about Grok: it does not live in a quiet productivity suite. It is built by xAI and woven into X, a network made for publishing. When an AI sits that close to a social network, the gap between "share" and "broadcast" can close without anyone noticing.
#Why the personal account is the problem
Consumer Grok, whether free, paid through SuperGrok or included in an X subscription, is not a business product. There is no data processing agreement and no admin console, so your organisation has no say over training, retention or sharing. The same goes for Grok inside X, which follows X's settings.
None of that makes the model bad. It makes the context risky. The same prompt typed into a governed business tool and into Grok on a personal account ends up in two different places, and only one of them has a share button that once turned chats into Google results. If you are the privacy officer who has to answer the Grok question, the useful question is not whether Grok is good, but which account people use it on and under which terms.
#What xAI does with your chats
xAI's privacy policy, last updated on 11 October 2026 and now issued in the name of SpaceXAI LLC, and its consumer FAQ fill in the rest:
- Training runs until you switch it off. Turning off "Improve the model" covers new conversations. Thumbs-up or thumbs-down feedback can still be used for training.
- Private Chat is not used for training and is deleted within 30 days, unless xAI has to keep it for legal, compliance or safety reasons. The same 30 days apply when you delete chats or your account.
- People at xAI can read chats. A limited number of authorised staff may review conversations to improve the model, investigate misuse or meet legal obligations.
- A share link is public. Anyone with the link can open the conversation, and a link posted publicly can be indexed.
- A data processing addendum exists only under the enterprise terms, where xAI also says it does not train on business data.
- Grok inside X falls under X's privacy policy and settings, not under xAI's. X's help page on Grok says your Grok interactions there may be used for training unless you opt out.
#Where the risk sits
Share becomes publish. This is the defining risk. Anyone who shared a work summary in 2025 could have made client data searchable. Shared ChatGPT links went through a similar episode weeks earlier.
Other people can read it. A conversation with Grok is not a private note: authorised xAI staff may review it, and on a personal account nobody in your organisation knows it exists.
Nothing to manage on a personal account. No retention set by an admin, no central offboarding when someone leaves. Meta AI has the same problem: an assistant inside a social app, used with a personal account.
#When a chat becomes a web page
Under the GDPR, putting an identifiable person's data into a consumer AI with no processor agreement is processing the employer cannot account for. If a "share" then turns that conversation into a public, indexable page, the problem moves from a controls gap to a data breach, one that must be assessed and, where there is risk to people, reported within 72 hours.
Fines reach €20 million or 4% of global annual turnover, whichever is higher. And the reputational risk is acute here: "our client's details were findable on Google because someone shared a Grok chat" is a sentence nobody wants to say out loud.
August and September 2024: training on EU posts paused. After an urgent High Court application by Ireland's Data Protection Commission, X agreed to stop using the public posts of EU and EEA users that it had processed between 7 May and 1 August 2024 for Grok training. In September it made that undertaking permanent. noyb filed complaints in nine EU countries, and in April 2025 the DPC opened an inquiry into whether training Grok on those posts was lawful.
August 2025: 370,000 conversations in Google. Forbes reported on 20 August 2025 that Grok's share button published conversations on Grok's website, where Google, Bing and DuckDuckGo could find them. Google had indexed more than 370,000.
2025: API keys in public code. Krebs on Security reported twice that a private xAI API key had ended up on GitHub: in May from an xAI developer, in July from a government employee. Both keys gave access to xAI's language models; neither report mentions exposed user chats. Something to weigh when you assess xAI as a supplier.
#Investigations and a court ruling in 2026
Since January 2026, most of the attention has gone to images: Grok was used on X to make sexual images of real people without their consent. Most of what followed is not about the privacy of your chats, but it does show how closely xAI and X are now being watched.
- 26 January, European Commission. A formal investigation under the Digital Services Act into whether X properly assessed and reduced the risks of bringing Grok into X in the EU, including manipulated sexual images.
- 3 February, UK ICO. An investigation into X and xAI over the processing of personal data to make sexual images without consent. The ICO has not reached a view yet; in October it said the investigation was ongoing.
- 17 February, Ireland's DPC. A GDPR inquiry into X Internet Unlimited Company over the same images, covering lawfulness, data protection by design and impact assessments.
- 26 March, Amsterdam District Court. At the request of the Offlimits foundation, the preliminary relief judge barred xAI (ECLI:NL:RBAMS:2026:3106) from using Grok to generate or spread sexual images that undress people living in the Netherlands without their explicit consent, and child sexual abuse material. X may not offer Grok within X while that continues. The penalty is €100,000 per day for each company, up to €10 million. It is not a general ban on Grok.
- 25 August, hidden instructions. Malwarebytes reported on research by Adversa AI: encrypted instructions hidden in a web page could make Grok, when asked to summarise that page, leak the user's name, approximate location, subscription tier and the prompts in the conversation. These are the researchers' own findings; they say they reported the flaw to xAI in June and that it was still open in August.
#The switches in X and Grok you need to know
1. Switch off training on grok.com and in the app On grok.com: Settings > Data > "Improve the Model". In the app: Settings > Data Controls > "Improve the model".
2. Switch off training on X Settings > Privacy & Safety > Data sharing and personalization > Grok & Third-party Collaborators > Data Sharing, then deselect the training option. A private account also keeps your posts out of training.
3. Use Private Chat when you do use Grok It keeps the conversation out of training and out of your history. Better than the default, not a guarantee.
4. Withdraw share links, and do not create them for work Treat any shared Grok link as a public page. Your existing links are listed at grok.com/share-links, where you can remove them.
5. Keep work data out of personal accounts The simplest control is not to use consumer Grok for work. If Grok is genuinely needed, use xAI's business terms with a data processing addendum.
Checked against xAI's and X's own policy and help pages on 20 October 2026. Menu labels change often.
#What a privacy switch does not cover
Turning off the training toggle changes what xAI does with your chats. It does not change the fact that Grok lives in a place where sharing is one tap from publishing, and it does not decide what an employee types into the box.
No setting retrieves a conversation that has already been shared and indexed. No opt-out undoes training on data that has already been used. The two risks that define Grok, a share feature that publishes and a person pasting sensitive context into a personal account, are beyond the reach of any configuration screen.
#The paste moment, and what you see there
BeeSensible checks the Grok prompt for sensitive data as you type. The text is sent to BeeSensible's EU detection service, where it is analysed in memory and discarded straight after detection. When sensitive content appears in a prompt, BeeSensible highlights it in the text, so the user sees exactly what they are about to share and can mask it with a placeholder such as [NAME1], or remove it, before sending.
By default the employee decides what happens with each highlight; your organisation chooses how warnings and masking are set up.
For Grok, this matters because the downstream risk is unusually public. If an employee is going to summarise a customer thread in Grok, the client name beside an account number is highlighted before it enters a conversation that might later be shared, and with Grok, shared can mean published. BeeSensible does not change X's defaults, and it does not replace a policy against work data in personal accounts or a recorded decision on whether Grok is allowed; those are the structural moves. It keeps the person aware of what is about to go in, while there is still time to keep it out.
Grok is engaging precisely because it lives where people already are: inside a social network. That is also why its mistakes become web pages. Keep work data out of personal accounts, never share sensitive chats and make sure the moment before a name is typed is a moment the person actually sees.
Grok
