CitadelSuite

Security

How CitadelGate keeps remote control under control

RMM tools are a favourite target for attackers because they can run commands on every managed computer. CitadelGate is being built on the assumption that it will be attacked, with layered controls between a request and a command running on a device.

Layers

Six layers between a click and a command

Tenant isolation

Provider, client, site, device group, device and user. Every query, ticket, command and session carries its client and is rechecked. Tests enforce isolation against a real PostgreSQL database.

Identity and MFA

MFA is required for every technician, administrator and owner. Client-scoped sessions last 15 minutes with no refresh token and are checked on every request. Revoking access blocks existing tokens immediately.

Separate permissions

Remote sessions, script runs, software deployment and role changes are separate permissions. Device credentials can't use human ticket, session or management endpoints.

Outbound-only agent

The Windows agent connects out to Gate over TLS. Enrolment is one-time and revocable, and agent updates are designed to be signed and verified before they install.

Ticket-bound actions

A PowerShell session or job is tied to a ticket number. No ticket, no command.

Complete audit trail

Each session or command records the actor, client, device, ticket, approval or consent state, script hash, timestamps, expiry, result and cancellation. Entries are hash-chained and append-only, so changes are detectable.

Automation rules

Rules for AI and automated fixes

AI remediation is built and in testing. These are the rules it follows.

  • Automatic runs are limited to signed library scripts labelled low risk.
  • Each client opts in, and can limit scripts or set maintenance windows.
  • AI-written PowerShell always needs technician approval.
  • Automatic fixes target only the device on the ticket.
  • Jobs are immutable once approved, time-limited and cancellable.
  • Bulk changes (planned) will start on a canary device and move in bounded waves.
  • The AI receives metadata and telemetry, not file contents.
  • Failed or ineffective fixes go to a person; there are no retry loops.
Attended and unattended

Consent comes first

Planned remote sessions come in two forms. Attended access asks the person at the computer to accept. Unattended access is only allowed for an enrolled client whose policy explicitly permits it, and is still logged against a ticket.

Hosting

Where Gate runs

The Gate API, worker and PostgreSQL database run in containers on infrastructure operated by Kemptechnology (PTY) Ltd, behind TLS at gate.citadelframe.com. Database backups are taken and verified before each schema change.

Frequently asked questions

How do I report a vulnerability?

Use the contact form, start your message with "Security", and include the details and steps to reproduce. We also publish a security.txt file at /.well-known/security.txt. Please don't test against accounts or data that aren't yours.

Does the agent open a port on the device?

No. The agent connects outward to Gate over TLS and identifies one enrolled device. The design rules out any unauthenticated inbound shell port.

Does CitadelGate hold any security certifications?

No. CitadelGate is in development and holds no security certifications. We describe the controls we are building and their status instead.

Help shape CitadelGate before launch

We are looking for a small number of South African MSPs and IT teams to pilot AI-assisted ticket remediation on test devices first. No cost, no commitment.

Join early access