Developers assume AI coding assistants only see what they explicitly open, but Cursor breaks that assumption at the very first prompt. A new analysis from ReclaimLLM confirms that Cursor receives the absolute path of your workspace folder before any tool execution occurs. This means the directory structure itself becomes part of the context window supplied to the model, potentially exposing sensitive client names and project details without a single file read.

The Path Is The Payload

In a staff reply dated November 12, 2026, Cursor confirmed that new chats ingest the opened folder's absolute path alongside project-specific directories located in the user's home folder. Consider a path like /Users/alex/Clients/ExampleCo/acquisition-review/payments-api. The model does not need to parse payments-api to know that ExampleCo is a client and acquisition-review is a sensitive business activity. The metadata is baked into the initial context, distinct from the permissions that govern later file access tools.

Privacy Mode Does Not Scrub Filenames

Teams often mistake Privacy Mode for a comprehensive data loss prevention (DLP) control. Cursor’s documentation states that Privacy Mode keeps code on your machine and prevents training usage, but it does not necessarily redact the path metadata sent during session initialization. Furthermore, .cursorignore controls tab and context indexing for code files, but it is not a DLP mechanism for directory names. If the path enters the context, exclusion rules for file contents do not hide the folder name from the model or the provider.

Audit Your Directory Hygiene

Security reviews must treat the initial prompt as a disclosure event. Before starting a session, inspect your workspace location on disk. If parent directories reveal client identities or confidential project phases, move the repository to a neutral folder or rename the parent directories. Additionally, check your installation settings for automatic indexing and model training toggles. Reviewer copies of sessions retain what was sent unless redacted before upload, so local redaction tools are critical for compliance.

Key Takeaways

  • Cursor receives absolute workspace paths in the initial context, potentially leaking client names from directory structures.
  • .cursorignore manages file indexing but does not function as a DLP control for path metadata.
  • Privacy Mode prevents code storage and training but does not guarantee path redaction in session logs.
  • Session evidence retained by providers includes initial context, requiring pre-upload redaction for compliance.

The Bottom Line

Stop treating directory names as invisible. If your folder structure screams 'Acquisition' or 'Client X,' the model already knows before you hit enter.