Enterprise AI requires more than model security.
AgentOS governs the entire chain from user intent to agent action. This page is the due-diligence front door: how the platform protects data, keeps humans in command, and produces evidence — stated as architecture facts, not certification claims.
Security posture, by construction.
Identity & access
Every actor — human, agent or service — holds a real identity; every request carries it. Multi-issuer trust integrates your IdP.
Tenant isolation
Logical boundaries across agents, data, credentials, workflows and observability, enforced at the platform layer.
Least privilege
Delegated authority narrows, never widens. Agents get the permissions their task needs — nothing else.
Encryption
TLS in transit; encrypted secrets at rest; mutual TLS available on internal agent communication.
Secrets management
Credentials are vaulted, scoped and injected at runtime — never stored in agent code or images. Agents use connections; they never hold keys.
Auditability
Decisions, invocations, approvals and resource consumption land in a queryable, tamper-evident trail.
Policy enforcement
Composite authorization (relationship + attribute engines), fail-closed. Unavailable policy = denied action.
Signed events
Outbound webhooks and event deliveries are HMAC-signed so your systems can verify origin and integrity.
Human oversight
Approval checkpoints and separation-of-duties gates on critical actions. People stay in command.
Your data stays yours — and access leaves evidence.
Tenant data ownership
Each tenant’s agents, credentials, documents, vectors and audit history live behind tenant boundaries. Deleting a tenant cascades: agents, data and identities go with it.
Access is always evidenced
No silent reads. Agent access to records flows through policy decisions carried out under a person’s delegated authority, and every access lands in the audit trail.
Credentials never leave the vault
LLM keys, database credentials and API secrets are stored encrypted, resolved inside the platform at call time, and are never returned to agents or users.
Data minimization by policy
Row- and column-level controls on structured data mean agents see the fields policy grants — not the table.
Governance is the product, not an add-on.
Most responsible-AI pages describe intentions. AgentOS enforces them at runtime — these are platform behaviors, not policy documents.
Humans stay in command
Critical actions pause at human checkpoints inside the workflow itself. The agent proposes; a person approves, rejects or escalates — enforced, resumable, recorded.
Authority is delegated, never assumed
Agents act on a person’s authority, cryptographically delegated hop by hop. It can be narrowed. Never widened.
Spend is an enforcement boundary
Budgets on teams, agents, people and models stop, pause or route work to a human by policy — not a dashboard you check after the bill arrives.
Every action attributes
Who asked, which agent acted, under whose authority, what it touched, what it cost, who approved it — one trail, built for evidence.
Built to carry your compliance program.
We do not list certifications we have not earned. What the platform provides today is the machinery compliance programs run on: complete audit evidence with delegation chains, separation-of-duties gates, approval records retained by policy, fail-closed authorization, and per-tenant isolation. As formal certifications are completed they will be published here — nowhere sooner.
Found something? Tell us directly.
Report suspected vulnerabilities to info@nirvanalogic.com with steps to reproduce.
How we handle reports
We acknowledge reports, keep you informed through remediation, and credit reporters who wish to be named. We ask that you avoid accessing other tenants’ data and give us reasonable time to fix before public disclosure.
Documents — deeper material, on request
Architecture overview, security whitepaper and questionnaire support are available under NDA. Request them at info@nirvanalogic.com.
Your infrastructure, your network boundaries, your keys.
Kubernetes-native
Runs on your clusters — cloud or on-premise. No dependency on our infrastructure to keep your agents governed.
Fails closed, degrades safely
If the policy engine can’t answer, the answer is no. Durable workflows resume across restarts — approvals survive infrastructure events.
Rollback-ready releases
Versioned images, declarative manifests and recorded release logs make every deployment reversible.