Security
Zero-trust segmentation for AI work.
Keep workspaces, identities, tools and approvals separated by what each job is allowed to do. The page shows the controls in place, the evidence available for review, and where each control stands.
The full security pack is sent on request, under NDA if you want one.
Last reviewed
Zero-trust segmentation
Every boundary has a reason.
A workspace should not see another workspace. An agent should not receive a tool it cannot use. A risky action should stop at a named person. That is the model we review with each customer and deployment.
| Review area | How we handle it | What gets confirmed |
|---|---|---|
| Independent assurance | Scope it | Tell us which framework or customer control set your review requires. We map the request to the deployment and confirm the evidence and status before contract. |
| Regulated data | Design it | Data boundaries, access, retention and human approvals are part of the workflow scope. If your review has a regulated-data requirement, name it before the build. |
| Security testing | Request it | Testing requirements and remediation evidence are evaluated with the deployment, risk tier and customer review—not assumed from a marketing page. |
| Continuity and service terms | Contract it | Backups, restore procedures, access boundaries and any uptime or response commitments are stated for the service and engagement you choose. |
| Customer requirements | Confirm it | We answer the questions that matter to your data, workflow and procurement process before the build starts. |
- Review area
- Independent assurance
- How we handle it
- Scope it
- What gets confirmed
- Tell us which framework or customer control set your review requires. We map the request to the deployment and confirm the evidence and status before contract.
- Review area
- Regulated data
- How we handle it
- Design it
- What gets confirmed
- Data boundaries, access, retention and human approvals are part of the workflow scope. If your review has a regulated-data requirement, name it before the build.
- Review area
- Security testing
- How we handle it
- Request it
- What gets confirmed
- Testing requirements and remediation evidence are evaluated with the deployment, risk tier and customer review—not assumed from a marketing page.
- Review area
- Continuity and service terms
- How we handle it
- Contract it
- What gets confirmed
- Backups, restore procedures, access boundaries and any uptime or response commitments are stated for the service and engagement you choose.
- Review area
- Customer requirements
- How we handle it
- Confirm it
- What gets confirmed
- We answer the questions that matter to your data, workflow and procurement process before the build starts.
Controls in place
Grouped by what a reviewer checks.
Each line says what the control does and whether it is complete. Detail on partial controls is in the security pack, which we send on request.
Identity and access
- Scoped, logged agent credentialsLive
- Credentials are scoped to the work, stored hashed and logged when used.
- Human access to productionLive
- Limited to a few named people and logged. Access reviews and retained evidence are part of the security work we scope with customers.
- Customer-managed identities and rotationPartial
- In progress. Today we issue and rotate agent identities for you.
Tenant isolation
- Workspace boundary enforced by the databasePartial
- Enforced in the database, not only in application code. Coverage is being extended and independently tested.
- Authorisation that fails closedLive
- A dedicated service decides who may act on what. If it cannot answer, the action is refused.
Agent controls
- Every agent action carries a risk tierPartial
- Actions an agent takes through the platform are classified before they run. The tier, not the model’s confidence, decides what happens next. Coverage of the remaining entry points is being completed.
- A hard ceiling on work done aloneLive
- Anything that publishes, pays, deploys or deletes is above what an agent may run alone. The platform enforces the ceiling, not the credential.
- Human-only approvalLive
- Approval routes accept a human actor. A model-generated confirmation, typed or spoken, is not a signature.
- Tools filtered before the model sees themLive
- A tool outside the allowed set is not offered and cannot be called.
- Memory advises, never authorisesLive
- Recalled context is labelled advisory. A remembered fact can inform a decision but never grant a permission.
Secrets
- Referenced, never inlinedLive
- Credentials live in a secret store and are injected by reference. None is carried into a prompt or a log.
- Rotated once exposedLive
- A credential that appears anywhere it should not is rotated, never reused.
Audit trail
- Every outcome writes an eventLive
- Completed, failed, denied or waiting for a person: each is an event under one correlation id.
Data and model providers
- Your data does not train modelsLive
- We do not train or fine-tune any model on your data.
- Provider list and retention termsPartial
- Each provider and its retention terms are listed in the security pack.
- ResidencyPartial
- One region today, on infrastructure we operate. Another region is a delivery engagement.
Not built yet
Four things a reviewer may ask for.
Downloadable evidence export
PlannedA bundle for any one run: who asked, which policy ran, the tools used, approvals, checks and costs. The events exist today, and we walk through them with you.
Single sign-on and directory provisioning
PlannedSign-on per tenant, with provisioning so removing someone from your identity provider removes them here. Designed, not built. Until then we manage accounts for you.
Workspace spending ceilings
PartialPer-run step and cost limits, loop detection and provider circuit breakers run today. The workspace-wide ceiling is in preview and being productised. Until then we set limits with you per run.
Checking business work with a second model
PartialOn supported runs, today software delivery, a different model family checks the work before a goal can close. Extending that to copy, customer replies and documents is being productised. Until then a person approves business work.
Pre-answered
The questions security teams send first.
Fifteen answers, a few sentences each. Send your own questionnaire and we will complete it, including the rows where the answer is no.
- Q-01Who are your subprocessors, and which model providers can see our data?
- The current list, with retention terms, is in the security pack. If a provider is unacceptable to you, say so early. Which model families run your work is decided with you.
- Q-02Do you train models on our data?
- No. We do not train or fine-tune any model on your data. Providers differ on retaining API traffic, so the pack lists their terms one by one.
- Q-03Where is our data stored, and can we choose the region?
- Production runs in one region today, on infrastructure we operate. Another region or your own environment is a delivery engagement, so raise it on the first call.
- Q-04How long do you keep our data, and what happens when we leave?
- Retention is set in the agreement. You can ask for an export or a deletion at any time, and we confirm in writing. There is no self-service retention control yet.
- Q-05Is our data encrypted?
- Traffic to the platform is served over TLS. The status of each encryption control is in the security pack, and we answer specific encryption questions in writing.
- Q-06Who at your company can see our data?
- A few named people, and access is logged. Credentials sit in a secret store and never enter a prompt or a log.
- Q-07What happens in an incident, and when would we hear?
- We tell you without undue delay, and within 72 hours of confirming an incident that affects your data. A written response procedure and a named owner exist. A public status page does not yet.
- Q-08Can security testing be part of the engagement?
- Yes, if your review requires it. Tell us the scope, environment and evidence your team needs, and we will confirm the testing plan and status before the work begins.
- Q-09What is your uptime commitment and continuity plan?
- We publish no availability figure and offer no SLA today. Backups run and a restore procedure is written. Continuity terms go in the agreement.
- Q-10What stops one tenant from reaching another tenant’s data?
- The boundary is enforced in the database, not only in application code. Coverage is being extended and independently tested, and the route-level inventory is in the pack.
- Q-11What do you log, and can we have it?
- Every action writes an event, whether it completes, fails, is denied or waits for a person, under one correlation id. Today we query it with you. A downloadable bundle per run is planned.
- Q-12How do you handle prompt injection?
- Retrieved content is treated as data, not instructions. More importantly, the model never holds the authority an injection wants: tools are policy-filtered, irreversible actions need a person, and memory never grants a permission.
- Q-13How is an agent identified, and what can it do on its own?
- Agent credentials are scoped, stored hashed and logged when used. The platform caps what an agent may do alone. Customer-managed identities with self-service rotation are in progress.
- Q-14Where does a person have to be involved?
- At every external, irreversible or regulated action: sending to a customer, publishing, spending, paying, deleting, deploying, filing. Approvals accept a human only, so an agent cannot approve its own work.
- Q-15Do your own people or agents act inside our workspace?
- Under managed operations, yes, with the same risk tiers as your agents and the same trail. If you need no standing third-party access, tell us early. It changes how we run the engagement.
Report a vulnerability
Found something? Tell us.
Email hello@vertixone.com with SECURITY in the subject. It reaches a person, not a queue. A dedicated address and a policy page are still to come.
What we commit to
- We acknowledge within two business days.
- We say what we plan to do and roughly when, then fix it or explain why not.
- We credit you by name if you want, and stay quiet if you do not.
- We will not pursue legal action over good-faith research within these limits.
What we ask
- Test against your own workspace and data.
- If you reach data that is not yours, stop at the proof and tell us straight away.
- No denial-of-service, load testing or social engineering.
- Give us a reasonable window before you publish.
What we do not offer
- No paid bounty programme today.
- No public hall of fame yet. Ask, and we will agree credit in writing first.
Get the detail this page leaves out.
The security pack has the control inventory with a status on every line, the provider list with retention terms, and the review work planned for your deployment. Ask for it and we send it to your reviewer.
See a proposed package before sharing contact details. A person then reviews your answers and sends a scope and figure.
