Cursor Tightens Grok Bot Training Rules After Employee Data Concerns

Cursor tightened access to Sand and moved to exclude employee-use data from future training after engineers raised concerns that confidential business information could reach data shared with SpaceXAI.

Sand was an internal version of Grok Bot that Cursor employees used for product testing and everyday work. In late July, engineers questioned whether activity on the tool could have brought confidential details about customers, suppliers or business partners into information sent to SpaceXAI for model training.

Cursor reviewed which users appeared in data it had already sent. It then required a Cursor email address to access Sand and changed future training practices to leave out employee-use data. Personal email use also complicated efforts to match employees to the data.

A Cursor lawyer wrote in a company Slack channel: “Legally, we arguably don’t have the rights to train on this data.”

Cursor’s access and training changes came as it worked with SpaceXAI on model development. On Apr. 21, Cursor announced plans to use SpaceXAI’s Colossus computing infrastructure to expand model training. On July 8, the company described Grok 4.5 as jointly trained with SpaceXAI on trillions of tokens of Cursor data reflecting user interactions with codebases and software tools.

Cursor’s description of Grok 4.5 training did not specify whether employee sessions with Sand were included.

Grok Bot entered beta on Aug. 11. Its agents are designed to work across a user’s apps and tools using dedicated computers, part of the product’s persistent cloud-computer setup.

SpaceX completed its acquisition of Cursor on Aug. 14. The all-stock transaction carried an implied equity value of $60 billion, with consideration in SpaceX shares. The review began weeks before the acquisition closed, while Cursor and SpaceXAI were already working together on training.

The episode highlights the challenge for companies deploying agents across workplace apps: keeping employee testing separate from data selected for model training.