Amid the wave of rapid development of large AI models and agents, the boundaries of internal corporate data security and privacy have once again attracted high attention. Recently, Cursor, the company responsible for development, encountered a data security storm, triggered by the possibility that confidential information generated during its internal testing and employees' daily use may have already flowed into SpaceX's AI division.
The focus of the incident points to Grok Bot (initially code-named Sand), an agent under SpaceX's AI division. As early as around month 6 of 2026, Cursor employees had begun using the tool in multiple work scenarios, and the company had collected operation records from various parties for model training. However, because some employees logged in with personal email addresses, the system had difficulty accurately identifying and stripping out the relevant data.
By late month 7 of 2026, the engineers in charge of development warned the company that this testing and daily-use data might contain highly confidential information about customers, suppliers, and others, and had already been transmitted to SpaceX's AI division. Facing potential compliance and security risks, Cursor immediately launched an internal investigation, though the specific conclusions of the investigation have not yet been made public.
To prevent similar problems from spreading further, Cursor subsequently took a series of remedial measures to make up for the lapse. The company not only required all employees to switch to company email addresses for logging in, but also made clear that employees' subsequent usage data would no longer be included in model training. Its internal legal team also explicitly pointed out that the relevant data may contain third-party trade secrets, and that the company itself may not necessarily have the legal authority to use it directly for model training.
Notably, the data security risk was not the only recent storm. Earlier this summer, Cursor had also experienced another customer information leak incident. At that time, its sales marketplace website built on Vercel had passwords written directly into the underlying code, putting some customer information of multiple companies at risk of exposure; fortunately, engineers discovered it in time and urgently took the website offline.
As Grok Bot gradually becomes a key core product in SpaceX's bid for the enterprise AI market, this series of incidents not only exposes management loopholes in the internal testing stage of cutting-edge tools, but also highlights the intricate security governance challenges among agents' high-privilege access, training data collection, and the protection of corporate secrets.