CONTROL FOR AI AGENTS

SIGMA ATG Enterprise

Designed to stop AI agents that try to act without identifying themselves or proving authorization.

SIGMA ATG Enterprise is designed to ask every agent to identify itself and prove what it is allowed to do before it reads data, changes records, or uses tools. If it cannot prove its identity or authorization, the operation stops. Sensitive actions can wait for a person to approve them.

Three conceptual SIGMA Command views: operations, agents, and approvals. Data is illustrative; these are not screenshots of deployed software.
Documented enterprise design
64 local tests of decision rules
192 cumulative local tests
Full integration to be validated in a controlled trial

THE PROBLEM

An agent should not inherit every permission of the person who signed in

Who is the person?

A person may sign in to work, but their session does not automatically let any agent read, change, or send everything they can access. Delegation must be specific and verifiable.

Which agent is trying to act?

Each agent must present a verifiable identity and be recognized locally. The planned console distinguishes internal, external, and unknown agents and associates them with a responsible owner. A valid signature alone does not grant access.

Is this action permitted?

The control compares the requested operation with the permission granted: resource, action, and limits. Permission to view a catalog does not allow price changes; permission to read a record does not allow export.

HOW IT WORKS

A decision before every effect

1 · Request

The agent requests an action

An external agent asks to view a record. The request states who the agent is, what it wants to do and which information it seeks.

2 · Verification

Identity and permission are checked

The system checks whether the organization recognizes the agent and whether it has current permission to view that record.

3 · Decision

A decision before any action

The request reaches the system only if it meets the rules. A sensitive action may wait for a person to approve it.

  • Allow: valid identity and permission.
  • Block: unknown agent or no permission.
  • Request approval: wait and check again.

CAPABILITIES

More control over what each agent does

Identified agents

The organization defines which agents it recognizes, who is responsible for them, and when they may act.

Per-operation permissions

Permission can be limited to a specific task, resource, and period. It is not open access to the whole system.

Human approval

A sensitive action can wait for a person’s approval before it continues.

Control before execution

Requests are checked before they reach the application or tool that would make the change.

Evidence and audit

The design records which agent asked to act, what was decided, and why, so each case can be reviewed.

SIGMA Command

Planned console for operations, permissions, decisions, and alerts with clear reasons.

REFERENCE ARCHITECTURE

Placed before the applications you want to protect

SIGMA ATG Enterprise sits on the route taken by requests to an organization’s portal, application, or tool. Before the system receives an instruction, the control point checks which agent sent it, whether the organization recognizes that agent, and whether this specific operation is authorized. A permitted read can continue, while an unauthorized change is stopped or held for human approval when the rules require it.

The proposal builds on the systems the organization already uses and defines where control belongs for each case. It also covers internal agents using tools managed by the company. A controlled trial checks connections, owners, limits, and the absence of alternate routes to the application. This validation is needed before claiming protection of a real enterprise environment.

Conceptual visualization of a control gate before the application and private decision services
Reference architecture. Actual services and connections require integration and validation.

For effective control, there must be no alternate route that lets an agent reach the application directly.

DEPLOYMENT MODES

Four ways to establish control

Protected inbound

Protects access to your portals and services: each agent action passes the control point before reaching the system.

Controlled outbound

Reviews what internal agents do when they use tools and connections the organization controls.

Legacy intermediation

Proposes supervised access to older systems. Protection depends on there being no direct route around the control.

Private operation

Covers private or restricted settings, subject to specific infrastructure and operational tests.

POSSIBLE APPLICATIONS

Where agents act, permission must be specific

Examples of possible uses. These are not customers or existing deployments.

Enterprises

An agent reads business data or prepares changes in internal systems with permission for each task.

Public administration

Portal and casework tasks with an institutional owner and traceability.

Healthcare

Bounded access to sensitive information by purpose and required approval.

Industry

Orders and operational support with system separation and allowed actions.

Energy

Automation on digital assets with stronger controls for sensitive changes.

Logistics

Agents querying inventory or coordinating flows with defined scope and receipts.

Commerce

External agents read or change business services only within their authorized scope.

Education

Institutional repositories with access scoped by identity and resource.

Start with one concrete, measurable case

1. Assessment

Choose an application, the agents that may use it, and the actions that matter.

2. Trial design

Agree on what should be allowed or blocked and how to measure the outcome.

3. Controlled integration

Connect the control to a test environment with a defined scope.

4. Adversarial tests

Try to act without permission, exceed limits, or bypass the control point.

5. Verdict

Review evidence, limits, and conditions before deciding the next step.

Conceptual visualization of five assessment and controlled PoC stages
Conceptual assessment path. Every step requires technical criteria and verifiable evidence.

LET'S DISCUSS A POC

Find the agent actions that need control

Together we select one system, the agents that use it, and a few important operations. We then design a controlled trial to check what is allowed, what is blocked, and what needs approval.

Request a SIGMA ATG assessment