Dark data hall with cool blue indicator lights

03Security and governance

Controls belong in the workflow. Matched to the action.

A workflow can take on more of the work when it has an identity, limited access, a rule for what it may do, and a person for the actions that matter. GreyShacks specifies those requirements with the assessment and the build. They are part of the workflow, scoped to its risk.

Scoped

To the risk of this use case

Limited

To the access the task needs

Approved

Where the action is consequential

Traceable

What was seen and done

01Who this is for

For the people who will be asked what the system was allowed to do.

Security leaders, CIOs, and the executives who own the process. A control model you can inspect against a specific workflow.

More of the work can be delegated when the bounds are part of the system.

The controls below are chosen for the risk of the use case. A routine classification does not need the same approval as a change to a customer promise or a financial record. Use the list before the first system is connected.

Layered architectural glass as a metaphor for permission boundaries

02Operating principles

Four decisions that keep governance usable.

03The control model

Ten things that have to be true.

Open a control. These are the bounds we design into a production workflow, scaled to what that workflow is allowed to do.

Control 01 / 10

Identity

Users have identities. Agents need identities too — separate, named, and reviewable. An agent is not “the platform.” It is an actor.

04Target state

What good looks like, in one room.

Night engineering floor
  1. 01You can list every agent and what it is allowed to do.
  2. 02You can reconstruct an action from a log, not from memory.
  3. 03High-impact writes require a named human.
  4. 04Shadow AI has a sanctioned alternative, not only a ban.
  5. 05Security can demote or disable an agent without a project.

05Systems we operate

Encryption is the same discipline as agent permissions.

When GreyShacks operates infrastructure, secrets are not stored in plaintext, and keys do not live next to the data.

01

AES-256-GCM

Sensitive connector credentials are encrypted before storage.

02

Envelope encryption

Google Cloud KMS. Database access alone is not enough to read secrets.

03

In transit and at rest

TLS 1.3 on the wire. Encryption on disk.

04

Your data stays yours

We do not train foundation models on private operational data.

IdentityPermissionsData boundariesHuman approvalsAuditabilityEvaluationMonitoringIdentityPermissionsData boundariesHuman approvalsAuditabilityEvaluationMonitoring

What would AI be allowed to do in your business?

Show us a process where AI still depends on people to gather information, make routine decisions, move between systems or finish the work.