Security
Know which controls the engine enforces.
Deeplinq combines technical controls with explicit deployment responsibilities. This page states the current boundary so security reviews can start from evidence instead of assumptions.
Security model
Defense in depth across the request lifecycle.
The strongest controls sit at the boundaries where identity, data, models, tools, money, and deferred work meet.
- 01
Bind tenant scope once
Resolve the organization at authentication and carry the immutable scope through authorization, storage, provider activity, jobs, and audit.
- 02
Deny before cost
Reject missing authority, unavailable controls, and insufficient balance before the request reaches a provider whenever the control contract requires it.
- 03
Minimize sensitive records
Exclude prompts, answers, evidence, secrets, and tool payloads from audit and tracing surfaces while retaining decision-grade metadata.
- 04
Make tampering detectable
Chain audit entries per organization and sign the chain head with a key held outside the database.
Control boundary
Current posture, without the procurement fog.
Engine-enforced means the application implements and tests the control. Shared control requires secure deployment configuration or organizational process.
| Control | Status | Current boundary |
|---|---|---|
| Tenant isolation | Engine-enforced | Forced PostgreSQL row-level security, role-separated database access, composite tenant keys, organization-scoped vector shards, and acceptance tests across two tenants. |
| Access control | Engine-enforced | Platform and organization roles, default-deny model and dataset grants, team membership, short-lived JWTs, and revocable API keys. |
| Audit integrity | Engine-enforced | Append-only per-organization SHA-256 chains with HMAC-signed head anchors, streaming verification, and NDJSON or CSV export. |
| Credential protection | Shared control | Connector credentials are sealed with versioned AES keys. Key custody, KMS protection, and rotation operations belong to the deployment. |
| Encryption in transit | Shared control | Provider connections use TLS. Public-edge TLS, HSTS, certificates, and internal network policy belong to the deployment. |
| Content safety | Configurable | Personal data, unsafe content, topic, and prompt-injection checks can screen prompts, outputs, tool arguments, and retrieved evidence. |
| Recovery | Shared control | Runbooks and destructive restore drills cover engine stores. Production backup scheduling, off-site copies, capacity, and organizational continuity remain deployment-owned. |
| Certification | Roadmap | The engineering control roadmap targets ISO 27001 first. Deeplinq does not claim a certification on this site. |
Operational assurance
Readiness before the first call.
The engine exposes platform and tenant-safe readiness views for chat, knowledge, agents, memory, and guardrails. Missing connectors, pricing, grants, credits, or sidecars become operator actions instead of hidden runtime surprises.
- Capability-level green, warning, and blocked states
- Structured logs and OpenTelemetry request correlation
- Runbooks for backup, billing reconciliation, key rotation, and vulnerability response
Known gaps
The roadmap remains visible.
The engine does not currently provide a general data-subject access request endpoint, breach notification system, complete retention policy across all records, AI fairness monitoring, or an organizational risk-management framework.
- Conversation and project memory deletion is implemented, general DSAR is not
- Audit data is a signal source, not a breach-detection and notification service
- Certification requires organizational governance and an independent audit
Next step
Review the control boundary with the people who built it.
Send your security questionnaire or architecture requirements. We will answer with current evidence and identify deployment-owned controls explicitly.