Cursor engineers investigated whether confidential customer and partner data had reached SpaceX’s AI unit for model training, during the internal testing of what became Grok Bot. Grace Kay at The Information reported it on October 15, behind a paywall, and the outlet re-shared it on X on October 18.
What the reporting establishes is an internal investigation and two fixes. What it does not establish is that any customer data was actually used for training. This page keeps those two things apart, sets the report next to what Cursor’s and SpaceXAI’s own documents say about training today, and lists what a team running Grok Bot should check.
On this page · 6 sectionsOpen
Cursor engineers investigated whether confidential customer and partner data reached SpaceX’s AI unit during Grok Bot testing, The Information reported October 15.
An internal lawyer warned that training on that data could create a “huge trust problem”; staff had not properly marked the data used for training.
Reported fixes: company email logins required for Sand, the internal Grok Bot, and employee-generated data barred from future training.
Grok Bot has no training policy of its own: training opt-out follows Cursor’s account settings, and with Privacy Mode on, customer data is not used for training.
Not established: that any customer or partner data actually reached a SpaceX training set. No finding has been made public.
§ 01What was reported
| Detail | Source |
|---|---|
| Cursor engineers investigated whether confidential customer and partner data was sent to SpaceX’s AI unit during Grok Bot testing | The Information on X, October 15 |
| An internal lawyer warned that training on that data could create a “huge trust problem” | The Information on X, October 15 |
| Staff “hadn’t properly marked data that they were using for training” | Grace Kay, in The Information’s clip, October 15 |
| The concern was raised in late July, in a Slack incident channel, about Sand, the internal name for Grok Bot | WebProNews, relaying The Information, October 16 |
| Some employees used Sand with personal email logins, which made their data harder to trace | WebProNews, relaying The Information |
| Fixes: company email logins required for Sand, and employee-generated data barred from future training | WebProNews, relaying The Information |
| SpaceX SI did not respond to The Information’s request for comment | WebProNews |
The shape of the problem is ordinary, which is why it matters. Cursor’s own employees were the testers, and they used Sand for real work. Real work at a coding company includes customers’ code and partners’ documents. When test usage feeds model training, whatever the testers touched can travel with it unless it is labelled and filtered first. Kay’s line names that failure in one sentence: the data had not been marked.
Two things are not in the public record. We found no statement from Cursor about what the investigation concluded, and no report that any customer or partner data was confirmed in a SpaceX training set. WebProNews put it plainly: no confirmed breach has surfaced.
§ 02The timeline
| Date | Event |
|---|---|
| Late July | Engineers raise the concern in a Slack incident channel |
| August 11 | Grok Bot opens in beta |
| October 12 | The X Subscription bundles Grok, Grok Bot, Cursor and X Premium |
| October 15, 13:00 UTC | The Information publishes the report |
| October 18, 18:00 UTC | The Information re-shares it on X |
The gap is the story’s second half. The worry was internal in July; the public heard about it in October, after Grok Bot had been in customers’ hands for two months and three days after Grok Bot became part of a single subscription with Cursor.
§ 03What the documents say about training
Grok Bot does not have its own training policy. SpaceXAI’s Grok Bot FAQ says it plainly: “Training opt-out follows the applicable Cursor account and privacy settings.” So the switch that decides whether your Bot’s work can train a model is Cursor’s.
- Privacy Mode on. Cursor’s data use page says “Customer Data will not be used for training by Cursor.” SpaceXAI’s security page repeats it for Grok Bot: “with Privacy Mode enabled, customer data is not used for training.”
- Privacy Mode off. The same Cursor page says “we may use and store codebase data, prompts, editor actions, code snippets, and other code data and actions to improve our AI features and train our models.”
- Legacy Privacy Mode. Not available for Grok Bot: “Grok Bot requires cloud data storage, so Legacy Privacy Mode is not supported.”
- The policy line. Cursor’s privacy policy, updated October 11, says “We do not use Inputs or Suggestions to train our models, or permit third parties to use them for training” unless the inputs are flagged for security review, reported as feedback, or covered by your explicit agreement.
Read against the report, the July problem sat in the gap these rules do not cover: employees testing their own company’s product, where the data was not marked as customer data in the first place. A setting protects data that is labelled. Kay’s sentence is about data that was not.
§ 04What to check if your team runs Grok Bot
- Confirm Privacy Mode at the team level. SpaceXAI’s security page says the team’s privacy mode governs a member while they are on the team. Check it is on, and who can turn it off.
- Require company logins. It is the fix Cursor applied to itself. Personal email logins made its own investigation harder.
- Know what flags data for review. Both Cursor’s data use page and SpaceXAI’s security page say providers may run abuse classifiers and store flagged data for investigation.
- Ask where tester data goes. If a vendor trains on internal usage, ask how it separates internal staff work from customer material, and how it labels each.
§ 05Why this matters beyond Cursor
AI teammates work on real company data all day: inboxes, codebases, customer files. That makes one question part of every buying decision for an AI employee: does my work train anyone’s model, and which switch controls it? At CellCog, none of the model providers we use trains on the data we send through their APIs, and their retention terms are listed in our privacy policy. Every vendor in this category should be able to answer the same question in one sentence.
§ 06What is still unknown
| Question | What would settle it |
|---|---|
| Did any customer or partner data reach a SpaceX training set? | A finding from Cursor or SpaceXAI |
| Were affected customers told? | A statement from Cursor |
| Was any model trained on that data retired or retrained? | A disclosure from SpaceXAI |
| Did the fixes cover external testers too? | Cursor’s own account of the changes |
Until one of those lands, the honest summary is short: an internal alarm, two fixes, and no public evidence either way on what already flowed.
Q1When did this happen?
Engineers raised the concern in a Slack incident channel in late July 2026. The Information reported it on October 15 and re-shared it on X on October 18.
Q2What is Sand?
As reported, Sand was the internal name for Grok Bot, which Cursor employees used for testing and real work before SpaceXAI opened the beta on August 11.
Q3Does Grok Bot train on my data?
SpaceXAI’s FAQ says training opt-out follows your Cursor account and privacy settings. With Privacy Mode on, Cursor and SpaceXAI both say customer data is not used for training.
Q4Did Cursor or SpaceXAI comment?
We found no public statement from Cursor on the outcome. WebProNews reported that SpaceX SI did not respond to The Information’s request for comment.
Q5What should a team running Grok Bot check?
That Privacy Mode is on at the team level, that members sign in with company accounts, and what the vendor does with data flagged by abuse classifiers.