Isolation belongs in the database.
Tenant-scoped tables enforce workspace boundaries with Postgres row-level security. Application code still checks access, but the database is the final gate.
Database-level tenant isolation, encrypted provider keys, signed webhooks, audit trails, and runtime guardrails for agents that talk to customers on phone and web.
Principles
The security model follows the surfaces your team actually touches: workspaces, provider keys, conversations, memory, tools, and webhooks.
Tenant-scoped tables enforce workspace boundaries with Postgres row-level security. Application code still checks access, but the database is the final gate.
Customer-supplied provider credentials are encrypted with per-credential data keys and only decrypted when a conversation needs them.
Assistant turns emit trace trees. Agent, tool, credential, and session changes keep actor and timestamp metadata where the platform mutates state.
Prompt-injection scrubbing, sensitive-memory rules, escalation rules, and topic cooldowns run inside the product flow instead of living only in internal policy.
Defense in depth
Identity, isolation, encryption, signatures, and observability work together so a voice-agent workflow can be reviewed, constrained, and traced.
Passwordless sign-in, short-lived sessions, scoped API keys, and workspace membership controls.
Workspace boundaries are enforced by the database, not only by request handlers.
TLS in motion. AES-256-GCM envelope encryption for provider credentials.
External events are signed, verified, replay-limited, and deduplicated.
Every turn is traceable. Sensitive payloads are truncated. Administrative changes are auditable.
Agent safety
Voice agents need controls that run where risk appears — on every assistant emission and every retrieval, not in a policy doc.
Per-agent keyword and regex filters inspect what the agent says before it goes out. Policies can block the turn, redirect to a canned reply, or escalate to a human.
PDFs, web pages, and uploaded text pass through prompt-injection scrubbing before chunks reach the retrieval layer, so adversarial documents lose their teeth on ingest.
Data control
The plumbing for user access requests, deletion, and retention is part of the platform — not gated to enterprise contracts.
A signed JSON bundle of memories, conversations, turns, open threads, and audit history. No pricing gate.
The forget workflow purges memory records and writes an audit row confirming the deletion path ran.
Hyponema does not train foundation models on tenant data. External model providers run under the terms attached to the keys you bring.
Recordings, transcripts, and observability spans each have their own retention windows, configured by workspace policy.
Compliance posture
No vague badge wall. The table separates what is available today from what is planned, and links to the legal artifacts already published.
Shared responsibility
Hyponema owns the infrastructure controls. Your team owns how agents are configured, who can access the workspace, which provider keys are connected, and how your end users are notified.
Hyponema owns
You own