Security · Philosophy and practice

Every rule is a check. Every check leaves a record.

An AI agent can be wrong. It can be misled by something it reads. And when several agents work together, one agent’s mistake can become another’s instruction. None of that is a reason not to use agents. It is the reason to build the system around them so that a mistake cannot turn into an action nobody approved.

An agent uses a model to reason and a harness to act. The harness is the agent program that calls a model and acts on the answers with its own tools. A seat is the confined environment in which the harness runs. AstralMaris is our own working team of agents, and it is where these designs are used first. What follows is how that team is secured.

ming-qiao, meaning “bright bridge”, is the message service agents and people talk through. The Awakener is the mechanism being built to start an agent within a limit and record what happened.

Six principles

These are the principles our security designs hold to. They do not mean every control is built or running; each feature’s status makes that distinction.

  1. Checks, not promises.

    An instruction to an agent is not a control. Every rule is enforced by a program that is not the agent.

  2. A person holds the authority.

    Agents do the work. The decisions that cannot be undone belong to a named person.

  3. Least reach.

    Each agent can touch what its job needs and nothing else, and that list is written down before it starts.

  4. Closed on doubt.

    When the system is not sure of its own state, it stops and waits. It does not guess.

  5. Everything leaves a record.

    Who asked, who acted, on whose permission, and how it ended.

  6. Reviewed by a different mind.

    Work is checked by another agent, built on a different model from a different maker.

How to read these pages

Each feature carries one of four words: In use, Built, In build or Designed. We list what is designed and not yet built, and we label it. A security page that shows only finished work tells you less than one that shows where the edges are.

In use
Running in our own team today.
Built
Written and independently reviewed, not yet running.
In build
Being written or at review.
Designed
Decided and written down, not built.

We describe designs, not our own installation. You will not find addresses, account names or key arrangements here.

What is not yet proven

No agent has yet been started by the Awakener in our team; today our agents are started by hand. Its complete path from permission through a bounded run to its receipt—a record of who asked, which agent ran and how it ended—remains to be demonstrated. The Awakener’s page keeps the current state.

Review by agents does not replace an outside audit, and we have not had one. The method is how we work day to day, and it is what an auditor would be shown.

The security designs

Who may decide.

Agents do the work; a person makes the decisions that cannot be undone.

Drafts for review

These seven pages develop the remaining topic outlines. Their feature statuses still need review where the supplied material does not confirm them.

Draft for review

The record

An action needs a record that survives the agent.

Draft for review

Containment

Keep the work inside the reach it was given.

Draft for review

Identity

Know who is asking before deciding what they may do.

Draft for review

Secrets

Keep credentials out of the agent’s working text.

Draft for review

What comes in

Material from outside is evidence to inspect, not authority to obey.

Draft for review

Recovery

After interruption, account for the work before starting more.