Security & governance

Built to clear a risk review.

Zahen treats security and governance as the product, not an add-on. Answers are access-filtered before retrieval runs, high-risk actions pause for human approval, credentials are brokered so agents never hold them, and every privileged action is written to an append-only audit log. It's documented here, ungated, for your security and procurement teams.

Grounded, access-filtered answers

  • Retrieval is filtered by your role and department BEFORE the search runs — there is no “filter after the fact”
  • Answers cite the approved sources they're drawn from; you can verify them in seconds
  • If the documents don't contain the answer, the assistant says so instead of guessing
  • Retrieved content is reference material, never instructions — built-in prompt-injection isolation

Human approval, by design

  • Risk lives on the tool, not the task — high-risk actions pause automatically
  • Nothing high-risk runs until a person with the right role decides
  • Five outcomes: approve, reject, edit & approve, request more info, escalate
  • No self-approval — you can't decide a request you raised, and this rule cannot be overridden

Append-only audit & observability

  • Every sign-in, question, document access, tool call, and approval is recorded
  • The log is append-only — entries cannot be edited or deleted after they're written
  • Filter by date, event, user, or department; export to CSV for review or your SIEM
  • Usage and estimated cost are visible by model and department

Identity & access control

  • Multi-factor authentication (TOTP) and modern password hashing
  • Role-based access control — roles are re-loaded from the system of record, not trusted from a token
  • Document-level access (public / employee / department / restricted)
  • De-privilege a user and their pending actions stop — permissions are re-checked at execution

Governed tools & credentials

  • Agents can only call tools an administrator has explicitly registered and enabled
  • Agents never hold credentials — the platform brokers each call from a secure store
  • Instant kill-switch: disable a tool or require approval on it without a code deploy
  • Tool inputs are schema-validated, and secrets / PII are redacted in the audit trail

Durable, resumable execution

  • Multi-step tasks are checkpointed at every step
  • If the system restarts mid-task, the agent resumes from its last checkpoint
  • A paused task holds its state until an approver decides — work is never silently lost
Human-in-the-loop

Approval is a decision, not just a gate.

When an agent proposes a high-risk action, the task pauses and an approver sees the proposed action, the risk level, and the agent's evidence and plan. They then choose one of five outcomes — every one of which is written to the audit log with who decided, what, and when.

1

Approve

The action proceeds exactly as planned; the task resumes automatically.

2

Reject

The action is blocked permanently; the task stops without performing it.

3

Edit & approve

Adjust the action (e.g. recipient or content); the task resumes using your edited version.

4

Request info

Send it back for clarification; it returns to the queue when answered.

5

Escalate

Hand it to someone else or a higher authority, with a per-request grant.

The rule that can't be bent

You can't approve a request you raised yourself. No self-approval — and it cannot be overridden.

Append-only audit trail

Reconstruct exactly what happened — line by line.

Every sign-in, question, retrieved source, approval, and tool call is recorded in its own transaction. Entries are never edited or deleted, and the whole trail filters by date, event, user, or department — and exports to CSV for review or your SIEM.

  • Who decided, what they chose, and when
  • Secrets and PII redacted in the record
  • Exportable to CSV for risk & compliance
Audit log · REQ-4827 Append-only
TimeActorEventDetail
09:41:02 AAgent Retrieved source refund-policy.pdf · clause 3
09:41:03 AAgent Proposed action billing.refund(AED 1,420)
09:42:18 RR. Haddad Approved with note · authorised
09:42:19 AAgent Executed tool billing.refund · 200 OK
Deployment & data handling

Dedicated and single-tenant — your data stays in your instance.

Zahen runs as a dedicated, single-tenant instance per customer. There's no shared multi-tenant SaaS boundary, so your documents, embeddings, and logs stay in your instance. Regional, private, and dedicated deployment options support regulated and government-adjacent organisations, with PDPL-aware, DIFC-aware, and ADGM-aware deployment support.

Editor's note: confirm the live regions, the model-training statement, encryption / key-management options, and any certifications with product + legal before launch. State only what is true today; mark the rest as roadmap.
Standards & framework alignment
  • OECD AI Principles — human agency and oversight
  • NIST AI Risk Management Framework
  • EU AI Act — human-oversight expectations
  • UAE AI guidance — Adoption Guideline, Maturity Self-Assessment, Ethics Guide
Bilingual & RTL

Zahen is bilingual English / Arabic and RTL-first. A language toggle flips the whole interface to right-to-left, in both the operator console and employee mode.

For procurement

Security answers without a form.

A security pack, architecture overview, and answers to common vendor-risk questions are available without a gate. Need something specific for your review?

Security & governance FAQ

Procurement-ready answers.

Do you support SSO?
Today, authentication is email/password with multi-factor authentication (TOTP). SSO is on the roadmap — talk to us about your identity-provider requirements.
Do you support Arabic?
Yes. Zahen is bilingual English / Arabic and RTL-first: a language toggle flips the entire interface to right-to-left, in both the operator console and employee mode.
Where do data and logs reside?
Zahen runs as a dedicated, single-tenant instance per customer — your documents, embeddings, and logs stay in your instance, not a shared SaaS. Regional, private, and dedicated deployment options are available; specifics are confirmed during scoping.
Is customer data used to train models?
State the precise position here once confirmed by the product and legal teams. [Needs the verified statement.]
How does this map to data-protection regimes?
Zahen provides deployment support aligned to your regional data-protection and compliance frameworks — including, in the UAE, PDPL-, DIFC-, and ADGM-aware deployment. This is deployment support, not a guarantee of compliance; data-protection obligations remain with your organisation and counsel.

See where governed agentic AI fits in your business.

Book a readiness workshop with our team. We'll map your highest-value, lowest-risk first workflow — no obligation to proceed.