Platform Capabilities
Security built for the AI-native stack
NeuralDLP processes text through a four-node pipeline executed in under 200ms. Every request is classified, evaluated against policies, enforced, and routed — with a full audit trail.
The Four-Node Pipeline
1. Classify
Data Classification
- 20+ PII entity types (SSN, email, phone, address, etc.)
- Credential detection (AWS keys, API tokens, private keys)
- Adversarial threat detection (prompt injection, jailbreaks)
- Risk scoring (0-100) with severity levels
2. Decide
Policy Evaluation
- Rule-based policy engine with priority ordering
- Per-entity actions: allow, redact, tokenize, block, escalate
- Tenant-configurable risk thresholds
- Auto-escalation for high-risk content
3. Enforce
Inline Enforcement
- Reversible tokenization (AES-256 encrypted storage)
- Redaction with strategies: mask, hash, remove, replace
- Blocking with structured reason and chain ID
- Human-in-the-loop escalation with callback support
4. Route
Provider Routing
- Route to alternative LLM providers based on risk
- Compliance-aware provider selection
- Configurable routing rules per tenant
- Fallback to default provider when no rules match
Enforcement Modes
Audit Only
Classify and log decisions without blocking or transforming. Use during rollout to understand data flows before enabling enforcement.
Enforce
Full pipeline execution. Tokenize, redact, or block based on policy. Transformed payloads are returned inline in the response.
High Risk Only
Enforce only when risk score exceeds the configured threshold. Low-risk content passes through with audit logging.
Reversible Tokenization
When enforcement mode tokenizes sensitive data, the LLM processes clean tokens. After the response, call /v1/detokenize to restore original values. Tokens persist for up to one year — safe for human-in-the-loop workflows with hours-long pauses.
Key Endpoints
| Method | Endpoint | Description |
|---|---|---|
| POST | /v1/orchestrate | Full pipeline: classify, decide, enforce, route |
| POST | /v1/orchestrate/batch | Batch process up to 50 texts |
| POST | /v1/detokenize | Restore tokenized values |
| POST | /v1/inspect/input | Standalone PII detection |
| POST | /v1/inspect/output | Output leakage detection |
| GET | /v1/security/summary | Dashboard metrics |
| GET | /v1/audit/logs | Audit trail with filters |
| GET | /v1/escalations | Escalation queue |
| POST | /v1/policies | Create security policies |
| POST | /v1/tenants/provision | Provision new tenant |