Image
Concepts

Exposing internal systems to agents without exposing your data

Connecting an agent to an internal system is easy. Connecting it so it sees only what the job needs, and cannot act beyond it, takes design. This page sets out that design: narrow views, separate accounts, read before write and a person on every consequential action.

The short answer

Short answer

You let agents work with internal systems safely by giving each one a narrow view built for its job: its own account, only the data the job needs, read before write, and a person's approval on every consequential action.

Anything an agent can see, it can repeat.

Today

Definition

Scoped agent access: access to an internal system given to an agent through its own account and a view limited to the data and actions one job needs, with consequential actions held for a person's approval.

What goes wrong when an agent gets the keys?

The quickest way to connect an agent is to give it an existing account, often one with broad access, and let it read whatever it needs. It works in the demo. Then the agent can see salary fields it never needed, customer records outside its job, and actions nobody meant it to take. An instruction hidden in an email or a document it reads can steer it toward any of them.

Who needs to settle this before anything is built?

The IT lead or security owner responsible for the system, the operations lead who wants the agent, and the developer who will connect it. The stakes rise with what the system holds: personal data, financial records, credentials, or actions that cannot be reversed.

How does scoped access compare with a quick connection?

Design choiceThe quick connectionScoped access
AccountA person's account, or a shared admin oneIts own account, named for the agent and the job
What it can seeThe whole systemA view with only the fields and records the job needs
What it can doWhatever the account allowsRead first; specific write actions only once justified
Risky actionsTaken directlyHeld for a named person's approval
RecordMixed in with the person's activityEvery call logged against the agent's account

How it works, step by step

Designing scoped access is a sequence the system owner runs before any connection is made.

01

Write down the job

State what the agent must do with the system, in a sentence. Everything else follows from that sentence.
02

List the data the job needs

List the fields and records the job requires and mark the sensitive ones. Leave out anything the job can do without, because, as OWASP's GenAI Security Project warns, LLMs, especially when embedded in applications, risk exposing sensitive data, proprietary algorithms, or confidential details through their output.
03

Build a view, not a door

Give the agent a view, report or interface that contains only that data, rather than direct access to the whole system.
04

Give it its own account

The agent gets its own account with read access to that view. Write actions are added one at a time, only when the job needs them.
05

Human gate: the system owner approves the access and every consequential action

The system owner approves the account, the view and each write action before it is enabled. Consequential actions, such as payments, deletions or changes to personal data, also need a person's approval each time they happen. OWASP traces excessive agency to the same three causes this step closes off: the root cause of excessive agency is typically one or more of excessive functionality, excessive permissions and excessive autonomy.
06

Log and review

Every call the agent makes is logged against its account, and the owner reviews the access on a schedule and whenever the job changes.
07

Remove what is no longer used

Access the agent has not used for a period you set is removed, the same way a person's would be.

Keeping data out of reach

Which rules keep the data out of reach?

One account per agent

Never a person's or a shared administrator account.

A view built for the job. Sensitive fields left out entirely.

Read before write

Write actions added only when needed, one at a time.

A person on consequential actions

The agent proposes; a named person approves.

Regular review. Access checked and trimmed like anyone else's.

Where this fits

Concept

Access design is a foundation for every workflow that reads an internal system, and the Platform page covers how it is built. Pick the jobs worth connecting with the readiness checklist. When several agents share systems, as in the agent-run operations back office, each still gets its own scoped access. Customers often ask how you control this, which is where security questionnaire automation comes in. The use case library shows the workflows these rules protect.

How does agent access differ from a new employee's access?

QuestionNew employeeAgent
How it learns what to doTraining and judgmentInstructions, which hostile input can imitate
What it does with what it seesUses discretionMay repeat it in output
How access should startThe role's standard accessNarrower still: one job's view
What stays the sameApprovals, reviews and removalApprovals, reviews and removal
AAOS Platform
How It Works

Questions, answered

Read-only still exposes data. Anything an agent can read, it can repeat in its output. A view limited to the job is safer than read access to the whole system.
AI Opportunity Audit · Agent access

Give agents a view, not the keys

Name one internal system and the agent someone wants on it, and see what scoped access would look like.