Safety and security · HelioStack

[ Safety and security ]

What stops it doing something stupid?

Rules, checks and approvals, never hoping the AI says no. All of it inside your own Azure.

[ Limits on the AI ]

What it can never do

Whatever the AI decides, these hold.

  • One signed step at a time

    Each action is checked against your rules, signed, valid for five minutes and tied to one task. When the task ends, nothing more can run.

  • No keys for the AI

    The AI models never hold Azure access. Only the broker beside your servers acts, and only with a signed step.

  • Nothing installed

    Azure's own Run Command does the work. No agent, no service, no open port on your servers.

[ Bigger changes ]

Many servers, never all at once

When a fix goes to many servers, it is rolled out in stages and stops at the first sign of trouble.

  1. 01A backup snapshot first
  2. 02One server first
  3. 03Its health checked
  4. 04Then the rest, a few at a time
  5. If anything looks wrongIt stops and undoes

Always needs a person: every new fix, changes to servers your rules mark sensitive, bigger production changes, and changes to the rules themselves.

[ Emergency stop ]

One button stops everything

Stop work on one server, a group of servers, or everywhere, at once. Lifting the stop releases nothing.

A pause for automatic work: coming later

[ Security and privacy ]

Private, keyless, and in your tenant

The questions a security team asks first.

01

Private by default

  • A private entrance and private endpoints.
  • No public endpoint for HelioStack.
  • No inbound connections to your satellites: they reach out.
02

No standing secrets

  • Managed identities and workload identity everywhere.
  • Local and key-based authentication turned off.
03

Where data lives

  • In your own Azure tenant.
  • AI models in the region you choose.
04

What the AI sees

  • Only what it needs for the task.
  • Secrets removed, and sensitive data pseudonymised where your rules say.
  • Unknown errors are researched online from Microsoft's cloud, never from your servers, with hostnames, addresses and user names removed first.

[ Everything recorded ]

A record you can check

  • Every request, decision and action

    Written to a tamper-evident ledger as it happens.

  • Kept for seven years

    Long enough for audit and incident review.

  • Cross-checked by Azure

    Azure's own activity log is an independent record to compare against.

NextTechnical →