Cached file contents encrypted with client-generated keys

The updated policy no longer explicitly discloses how Cursor handles plaintext code and metadata during codebase indexing. Previously, the policy stated that plaintext code ceases to exist after each request and that embeddings and metadata may be stored in the database. The removal of this language creates ambiguity about current data handling practices for users who index their codebases.

The updated policy clarifies that Cursor maintains zero data retention agreements with all AI model providers and customer data will not be used for training by Cursor. However, the policy now explicitly discloses that model providers may run risk classifiers to detect policy violations, and if your prompts or conversations trigger abuse detectors, your data may be stored for investigation and deleted according to the provider's retention policies. The policy removed the previous blanket statement that code would never be trained on by Cursor or third parties, replacing it with more specific disclosure of abuse detection practices. You can review OpenAI and Anthropic's documentation directly for details on their specific retention policies.