Data Leakage Prevention (DLP)

The foundational capability to automatically locate and categorize sensitive data across the entire digital estate. This process scans structured databases and unstructured repositories like file shares and cloud storage.

• Exact Data Matching (EDM): Fingerprints structured data like customer records to detect full or partial matches.
• Indexed Document Matching (IDM): Creates a fingerprint of a sensitive document to detect derivative works or leaked sections.
• Content Classification: Uses regular expressions, keywords, and machine learning to identify PII, PCI, and PHI.
• Contextual Analysis: Goes beyond keywords to analyze the surrounding text and metadata for more accurate classification.

The engine that inspects data in motion and at rest against a defined rule set. Policies are granular and context-aware, evaluating not just the data content but also the user, destination, and action.

• Endpoint Monitoring: Tracks data transfers to USB drives, printers, and local applications.
• Network Traffic Analysis: Inspects HTTP, FTP, SMTP, and cloud application traffic for policy violations.
• User and Entity Behavior Analytics (UEBA): Establishes a baseline of normal user activity to detect anomalous data access or exfiltration attempts that deviate from the norm.
• Policy Templates: Pre-configured rules for common regulations like GDPR, HIPAA, and PCI DSS accelerate deployment.

The ability to execute a pre-defined, automated response when a policy violation is detected, minimizing the window of exposure and reducing manual workload for security teams.

• Blocking: Prevents the action entirely, such as stopping a file upload to a personal cloud drive.
• Encryption: Automatically applies encryption to sensitive data being transferred to a removable device.
• Quarantine: Moves a sensitive file found on an unauthorized share to a secure, access-controlled location.
• Just-in-Time User Coaching: Displays a pop-up warning to the user explaining the policy violation and allowing them to self-remediate or provide a business justification.

A centralized management console that provides deep forensic auditing and compliance reporting. This capability transforms raw incident data into actionable intelligence for stakeholders from security operations to the C-suite.

• Pre-built Compliance Dashboards: Map DLP incidents directly to specific regulatory controls (e.g., GDPR Art. 32, PCI DSS Req. 3) for auditor-ready reports.
• Forensic Audit Trails: Provide a tamper-proof, chronological record of every policy violation, including the user, data involved, destination, and action taken.
• Risk Posture Analysis: Aggregates incident data to identify high-risk users, departments, and data types, enabling proactive security improvements.
• SIEM/SOAR Integration: Forwards rich, contextual DLP alerts to a security information and event management platform for cross-tool correlation and advanced orchestration.

A security mechanism that restricts access to entire documents based on a user's identity or group membership. In a RAG pipeline, this prevents the retriever from pulling chunks from documents the user isn't authorized to see.

• Enforced at the index partition or metadata filter level
• Uses ACLs to map user identities to document IDs
• Critical for multi-tenant knowledge bases where different clients share the same vector store
• Example: A legal AI assistant must never ground an answer on a contract from a different client, even if semantically relevant

The process of filtering search or retrieval results to exclude content that the querying user does not have permission to see. This is the runtime enforcement of access policies.

• Can occur pre-query (filtering the index before search) or post-query (removing unauthorized results after retrieval)
• Pre-query filtering is more efficient but requires permission metadata to be indexed alongside vectors
• Post-query filtering is simpler to implement but can lead to empty result sets if top-k results are all unauthorized
• Essential for compliance with data residency and confidentiality requirements

An access control paradigm that evaluates attributes of the user, resource, and environment against a set of policies to grant or deny access dynamically. ABAC is the modern standard for fine-grained DLP in retrieval systems.

• Attributes include: user department, document classification, time of day, device posture
• Policies are expressed as Boolean rules: grant if user.clearance >= document.sensitivity
• Enables context-aware retrieval that adapts permissions based on session context
• More scalable than RBAC for large, dynamic document corpora

An access control technique where a user's security permissions are applied as a filter before a query is executed against an index. This ensures only authorized documents are scored and retrieved.

• Implemented via metadata predicates in vector databases (e.g., Pinecone, Weaviate, Milvus)
• Requires an authorization sidecar that resolves user identity to a list of permitted document IDs or labels
• Dramatically reduces the risk of data leakage because unauthorized content never enters the retrieval pipeline
• Trade-off: Filter cardinality can impact query latency if permission sets are highly granular

The enforcement of data access controls during the retrieval phase of a RAG pipeline. This ensures the language model only grounds its answers on documents the user is permitted to see.

• Combines identity propagation from the application layer through to the vector store
• Requires the authorization context to survive query rewriting and hypothetical document embeddings
• Failure mode: A model synthesizes an answer from a document the retriever should have filtered out
• Emerging best practice: Decoupled authorization where a dedicated service issues a cryptographically signed capability token scoped to permitted resources

A metadata label applied to a data asset that indicates its sensitivity level and required security controls. Classification tags are the atomic unit of automated DLP enforcement in retrieval pipelines.

• Common taxonomies: Public, Internal, Confidential, Restricted
• Tags are applied during sensitive data discovery using regex patterns, ML classifiers, or manual labeling
• Vector databases use these tags as filterable metadata fields during pre-retrieval filtering
• Example: A chunk tagged PII:True is automatically excluded from retrieval for users without data protection clearance