Security & Governance

The strongest control is architecture.

Most AI security conversations are about trusting a provider’s promises. ZenithAI changes the conversation: when the platform, the models and the knowledge all live inside your own security boundary, the decisive control is structural — and the remaining controls are built into the product.

Sec. 01 Data residency

Your data stays where you put it.

The core platform — chat, models, embeddings, retrieval, computation, document generation and OCR — processes organisational data entirely on infrastructure you control. There is no external model API behind it and no third-party processing of your prompts, documents or knowledge by the core platform.

Alignment with Indian data-protection law. Processing personal and organisational data inside your own infrastructure directly supports data-minimisation, purpose-limitation and localisation objectives — including obligations that apply to data fiduciaries under India’s Digital Personal Data Protection Act, 2023. Your legal team defines the policy; the architecture makes it enforceable.

The decisive line

The highest-stakes control in the system is the consent boundary for private-document egress. It is the one line where a mistake means document content leaves the building, so it is engineered to fail closed. A document or workspace chat reaches the internet only after the user allows it for that chat; the search words can then include terms from the documents, so every query and page opened is logged, and labelled secrets such as passwords and API keys are screened out. Any consequential send, an email or a call, waits for a person’s explicit confirmation. By default nothing about your documents has to leave the infrastructure you control; when an optional connector does reach out, that activity is logged and screened for sensitive content. Every other control on this page is defence in depth around that single boundary.

Sec. 02 Product controls

Controls built into the platform.

SC-01AuthenticationSign-in verified with time-based one-time passwords; optional persistent sessions bounded by an administrator-set expiry; tokens stored hashed, never plain.Enforced
SC-02Role separationAdministrative functions require an authenticated administrator account — API keys and internal calls can never satisfy an admin check.Enforced
SC-03Workspace isolationPrivate to creator until deliberately published; cross-user requests are refused without confirming a workspace even exists.Enforced
SC-04Fail-closed privileged auditAdministrative access to user content is written to a dedicated audit ledger first; if the audit write fails, the content is withheld.Enforced
SC-05Egress visibilityOutbound activity from optional connectors — email sends, web requests — is logged, with secret values such as passwords and API keys redacted.Enforced
SC-06Hardened executionAI-generated analysis code runs in a locked-down sandbox: restricted language surface, resource limits, no network, no filesystem reach.Enforced
SC-07Secret hygieneKeys and tokens stored hashed with atomic, corruption-safe writes; a redaction layer screens replies, chat history and outbound search queries for labelled secrets such as passwords and API keys.Enforced
SC-08API governanceManaged keys with per-key usage tracking, issued and revoked by administrators; transport security enforced at your gateway with TLS.Enforced
SC-09Human-consent gatesConsequential actions — sending an email, placing a call — require explicit, contemporaneous confirmation. The assistant proposes; a person decides.Enforced
SC-10Internet you can seeThe message box shows whether the internet can be used for the next message (since 21-10-2026 as a globe next to Send; installer 4.6.0 has a Search button and an Internet chip). A chat with private documents goes online only when the user allows it for that chat or types #search; administrators can switch internet off for any person. How it works.Enforced
SC-11Attachment retentionWhen the administrator switches keeping on, files attached in chat are kept on your server, for 90 days by default or the number of days the administrator sets, and only the person who attached them can open or download them. Deleting a chat or an account removes its files, and the server administrator sees and manages the space they use. Live on ZenithAI since 21-10-2026; not in installer 4.6.0.Configurable

Sec. 03 Shared responsibility

Clear lines, honestly drawn.

Because ZenithAI runs in your environment, security is a partnership. We are explicit about which controls the product enforces, which the deployment configures, and which remain yours — so your CISO evaluates a real architecture, not a slogan.

LayerResponsibility
Product controlsEnforced by ZenithAI: authentication, isolation, auditing, sandboxing, consent gates, secret hygiene
Deployment controlsConfigured per engagement with our team: network placement, TLS termination, module enablement, retention, backup
Customer environmentOperated by you: physical security, host OS and network hardening, identity lifecycle, disk encryption policy
Optional connectorsJointly governed: each outward-reaching module is enabled deliberately, scoped and logged

Sec. 04 Governance

Administration with oversight built in.

User overview, API-key management, model and system health, and audited access to user activity when duty requires it — every such access itself recorded in the privileged-access ledger.

Enterprise-defined policy. Which modules are enabled, which workspaces are shared, how long sessions persist, what leaves the network — these are your decisions, expressed as configuration, not requests to a vendor.

ZenithAI does not currently claim third-party security certifications for the product itself; security posture is demonstrated through architecture review and deployment-specific assessment with your team.

Questions

Security, answered.

Where does my data go when I use ZenithAI?

Chat, models, embeddings, retrieval, computation, document generation and OCR all run on infrastructure you control. There is no external model API behind the core platform, and the core platform sends your prompts and documents to no third party.

Can ZenithAI send anything to the internet?

Only through optional connectors. A document chat reaches the internet only after the user allows it for that chat. Because the search words can include terms from the documents, every query and every page opened is logged, and labelled secrets such as passwords and API keys are screened out. Sending an email or placing a call always waits for a person to confirm.

Does ZenithAI help with the DPDP Act?

It supports obligations under India's Digital Personal Data Protection Act, 2023, such as data minimisation and keeping data inside your own boundary. Your legal team sets the policy, and the architecture makes it enforceable. We do not currently claim any third-party certification.

How is access controlled?

Sign-in uses time-based one-time passwords. Administrative functions need an authenticated administrator account, and API keys cannot pass an admin check. A workspace is private to its creator until someone deliberately publishes it.

What happens if the audit log fails?

The content is withheld. Administrator access to user content is written to the audit ledger first, and if that write fails, the content is not shown.

Is AI-generated analysis code safe to run?

It runs in a locked-down sandbox with a restricted language, resource limits, no network and no access to the filesystem.

The next step

Bring your CISO. We’ll bring the architecture.

We welcome security review as part of every evaluation — walkthrough, data-flow diagrams and control mapping against your framework.