Security at Axari. Every control, and how it works.

Axari carries security work across your most sensitive systems. Here is what holds each Twin in bounds, how we protect your data, and where to get the evidence for your review.

Agent controls. What a Twin can reach, and what holds it back.

The questions an agent review asks, one row each. The full detail is section 12 of our data security practices.

Agent controls: what each one covers and how it works
ControlHow it works
ReachOnly what its person can.Each Twin runs in its own VM, on its owner’s own authorizations, scoped per user and per tenant. Connect only what each responsibility needs.
CredentialsNever reach the Twin.A Twin acts through Axari’s gateway, so credentials never enter its environment or reach the model. They’re tested on save, stored encrypted and never logged, and each person can view and revoke their integration grants at any time.
ApprovalRisky actions wait for the owner.Every tool call is marked as a read or a write, and anything unclear counts as a write. By default, sending outside, spending money, changing access and deleting data wait for the owner to approve, in Slack or Teams. Owners can let low-risk actions, like simple replies, run on their own. Approving a pull request, or anything involving secrets, always waits.
RecordEvery step is logged.Sessions, actions and approvals go to an activity log: which Twin, in which tool, under whose approval, with the evidence behind each result.
VisibilityA Twin’s work belongs to its owner.Workspace admins and other users can’t see it. Shared context is limited to what each person is allowed to see, and personal memory is never pooled across customers.
Model trafficYour key, your provider account.With your own model key, agent inference goes to your provider. Platform calls such as classification use Axari’s keys. Your key is stored encrypted, never logged, and deleted on termination or on request.
MonitoringAgent behavior is watched in production.LLM observability runs on our own stack. Session telemetry is summarized and redacted on arrival, and protected like customer data.
TestingTested against the OWASP Top 10 for LLMs.Including prompt injection, excessive agency and sensitive data leaks, the risks specific to AI agents.
Off switchRevoke any grant, any time.Uninstalling Axari revokes every stored grant, including model keys.

Isolated and encrypted. How we protect your data.

Axari has completed a SOC 2 Type 1 examination of the Security Trust Services Criteria, by Prescient Assurance. Below is a summary of our published data security practices; the full document is the reference.

Data protection

Encrypted at rest and in transit, isolated per tenant.

Encryption

AES-256 class encryption at rest for confidential data and TLS 1.2+ in transit, with managed, rotated keys.

Tenant isolation

Multi-tenant isolation at the application and data layers. Any cross-tenant support access requires explicit authorization and is logged.

Retention and deletion

After termination or uninstall, customer data is kept 30 days for export, then deleted, and stored OAuth grants are revoked.

Backups

Regular encrypted backups of critical systems, with restoration procedures tested periodically.

How we run Axari

Controls in layers, from code to people.

Identity and access

SSO and OAuth, MFA for administrative access, role-based access, and access reviewed on hire, role change and departure.

Secure development

Peer-reviewed changes, automated security checks in CI/CD, and secrets kept in secret managers, never in source.

Sub-processors

Sub-processors handling sensitive data are security reviewed, with DPAs where required and incident notification in contract.

Incident response

Incidents involving customer data are notified to affected customers without undue delay, with root cause and corrective actions tracked.

Common questions

How does a Twin handle prompt injection?

It doesn’t take orders from what it reads. Emails, tickets, documents and scanner findings are material to work on. A Twin takes direction only from its owner, in their own Slack or Teams chat, and from work its owner has set up to run.

No AI model can be guaranteed never to be misled by crafted text, so a Twin’s safety doesn’t rest on that. It rests on limits to what a Twin can do: every tool call is marked as a read or a write, risky actions wait for the owner, credentials never enter the Twin’s environment, and each Twin runs in its own VM. Agent controls has the detail.

What does Axari collect?

Your account and configuration, the integration grants you connect, and the work your Twins do: prompts, context, memory and outputs. Session telemetry is summarized and redacted when it arrives.

Where does it live?

In Axari’s hosted service, encrypted at rest (AES-256 class) and in transit (TLS 1.2+), isolated per tenant and per user. Processing locations and sub-processors are documented; ask for them as part of your review.

How long is it kept?

It depends on the type of data. After termination or uninstall, customer data is kept 30 days for export, then deleted, and stored grants, including your model keys, are revoked. Backups expire on their own schedule.

Is it used to train models?

No. Your data is never used to train models.

Doing a security review? Start here.

Start in the Trust Center for documents and policies, or email us for the SOC 2 report and anything else your review needs.

This page was last updated on 13 November 2026.