AI Data Security

AI applications often move the same information through several representations and services. A record may appear in a training file, an embedding index, a prompt trace and an answer cache, each with different access controls and retention behavior. Data security treats these as parts of one information flow rather than assuming the model endpoint is the only boundary. Confidentiality concerns who can learn the information; integrity concerns whether it can be altered without authorization. Privacy adds questions about people and permitted processing, while security also covers credentials, commercial secrets and other protected assets.

The practitioner inventories sensitive data and traces its paths through preprocessing, inference, retrieval and monitoring. They configure least-privilege identities, protect secrets outside prompts, restrict outbound destinations and decide which fields should be removed before external processing. A data-flow review should produce an access matrix, retention settings and tests for disclosure through errors, logs and tool responses. When a model provider is involved, the team verifies the actual service configuration and contractual data handling rather than inferring them from the word enterprise.

Redaction detectors miss unusual identifiers, and encryption does not help when an authorized application sends decrypted secrets to the wrong destination. Access policies must cover derived data and backups as well as originals. Security reviews should distinguish prevented disclosure from merely undetected disclosure, and verify negative paths. A clean sample of model answers does not prove that training membership, cached context or agent tools cannot reveal protected information.