Decide what code context you are authorised and comfortable disclosing before using Cursor's AI features. Do not assume that working in an editor means prompts and repository context stay on your machine.
Understand the documented data flow
Cursor's privacy documentation says AI features send prompts and code context to model providers. With Privacy Mode enabled, Cursor says code is not used for training by Cursor or other AI model providers.
These statements answer different questions. A training restriction is not a statement that no transmission occurs. Your review should separately ask what information may leave your environment, for what purpose, and under whose approval. These are operational review questions, not legal conclusions.
Build a repository disclosure map
Use four classifications as editorial guidance:
- Public and approved: Material already public that you have permission to use for this task.
- Internal and reviewable: Non-public logic requiring a named owner's approval before disclosure.
- Excluded: Secrets, credentials, customer records, private messages, and production data.
- Unknown: Files whose ownership or contents have not been reviewed.
For each folder or proposed snippet, record its classification, owner, reason it is needed, and a safer substitute. Public availability alone should not replace a permission review. Keep unknown material out of the proposed context until someone resolves its classification.
This map is a decision document, not a claim about Cursor's automatic selection or exclusion controls.
Try a hypothetical disclosure exercise
Imagine a Bengaluru developer preparing a demo repository with three files: a public sample formatter, an invented discount function, and a dummy environment template containing placeholders only.
The task is to explain a rounding defect in the formatter. The minimum proposed context is the formatter plus two invented inputs: ₹199.50 and ₹200.49, with explicitly stated expected rounding behaviour.
Evaluate each additional file by asking whether it changes the diagnosis. The discount function is unnecessary; the environment template is also unnecessary. Exclude both rather than sending a broader bundle “just in case”. Never insert real credentials into the example.
Stop at unresolved disclosure
Stop if completing the task requires private material or permission remains unclear. Next, construct a synthetic reproduction or ask the responsible owner for a review of the proposed context. Do not treat a marketplace purchase as approval to disclose organisational code.
If access options matter after that review, check the coding catalog, verify access format, and save offer terms before buying. Marketplace account/code/token offers are not proof of official subscription equivalence.
Sources and scope
Source material checked 29 October 2026; features/policies may change. Cursor privacy and data governance establishes provider transmission and Cursor's stated Privacy Mode training restriction. The classification map and minimal-context exercise are editorial advice, not guarantees about context filtering or marketplace privacy terms.