Skip to content
InovagioInovagio

Series · How we solve it

Is it safe to give AI a login?

Work that crosses systems needs a login, and a login is what a security team guards hardest. These are the five controls that make a Virtual Employee safe to credential, how the kill switch works, and the one number a security team should watch.

· Published · Updated · 8 min read

On this page

The ticket reaches the security team on a Tuesday afternoon. Finance operations wants a service account for "the AI" so it can pull invoices and update vendor records in the finance system. The request names no owner and no end date, and it asks for read and write access to the whole system. Everything the security team has learned says to send it back. A shared account with broad rights, a password in a configuration file and a program that reads documents from outside the company: that is how incident reports begin.

Yet the work in question crosses several systems, and work that crosses systems needs a login. The useful question about AI agent identity and access is whether that login can behave like an employee's: issued to one named principal, scoped to the job, watched while it works, and revoked the moment it isn't needed.

The industry calls this software AI agents or AI employees. Ours are Virtual Employees, and what makes one a worker rather than a chat window is covered elsewhere. This piece is about what it takes to give one a login, and it is written for the person who has to sign the access request.

Two kinds of Virtual Employee, two levels of access

#

Not every Virtual Employee needs an account on your network. Many work from a queue. Requests arrive by email or in a shared folder, and the Virtual Employee reads and writes only through approved connectors. It has no account in your directory at all. That is the right place to start, and for plenty of queues it is where the work stays.

The second kind works on your network, because the job needs a system no connector reaches. That Virtual Employee has an account in your directory, and a desktop of its own when the job needs one. Every session is recorded, and it works in shadow mode, with your team still deciding, before it takes on live work. Everything below applies to both kinds, but it matters most for the second.

What differsWorks from a queueWorks on your network
Account in your directoryNoneIts own, with least privilege
How it reaches your systemsApproved connectorsThrough the gateway
Sessions recordedEvery task leaves a receiptEvery session, plus the receipt
Shadow mode before live workYesYes

Five controls for AI agent identity and access

#

Your approved rules tell a Virtual Employee what the right work is. Four other controls make sure it can't do anything else. Each is described below by what the system does.

The five controls for AI agent identity and access: identity, reach, checked plans, approval and your rules, each in one line
Your rules say what the right work is. The other four make sure it can't do anything else.

Identity

#

Every Virtual Employee is a real principal with its own roles and group memberships, and there are no shared accounts. For a Virtual Employee on your network, Auth, our identity service, federates the identity to your identity provider. Your conditional access and privileged-access rules then apply to it as they do to a person, and so do your access reviews. Extra rights are granted just in time, and each grant is audited. Disable the account and the work stops.

Reach

#
The life of one step's credentials: the gateway issues credentials when the step starts, the step reaches only the destinations it needs with each request logged, and the credentials are revoked when it ends
Credentials never appear in a prompt and are never written to disk.

Every call to one of your systems goes through a gateway that allows only the destinations that Virtual Employee needs and logs each request. There is no open internet and no moving sideways from one system to the next.

The gateway also holds the credentials. Each set is short-lived and scoped to one step: issued when the step starts, revoked when it ends, and never shown in a prompt or written to disk.

Checked plans

#

The Virtual Employee proposes a plan. Before anything runs, the plan is checked against its role and your approved rulebook, and against the data it has been granted. The checked plan then runs inside a sandbox step by step, exactly as approved. This is the control that answers prompt injection. A poisoned document can change what the Virtual Employee proposes, not what the checks allow. Screening prompts, documents and tool calls for injection is a second control, run by Guard, and business rules and AI guardrails explains why regulated work needs both.

Approval

#

Steps your policy marks as approval-required wait for a named person. Work the checks can't confirm goes to a Needs a person queue, and dual control applies wherever your rules require two approvers.

Your rules

#

Approved rules are hard gates, so only work that fits them goes ahead. A correction becomes a new, versioned rule with an approver, which we describe in a mistake your team fixes once stays fixed.

Together, the five controls tie every action to one identity and one checked plan.

Stopping one Virtual Employee, a department or everything at once

#

A security team will ask how fast it can stop the thing. The kill switch in Archon, the control plane, works for one Virtual Employee, for a department or for everything at once, and Guard enforces it. It sits on the same screen that shows what each one is doing.

Stopping AI agents at three levels: one Virtual Employee, a department or everything at once, with each stop recorded
Every stop is recorded: who, when, why, and what the work was doing.

That screen shows, live, which Virtual Employees are running, the step each is on, what it is touching, the cost so far and every attempt the policy check blocked. From it, an operator can pause or stop work at any of those three levels and start it again. The same screen replays any session and sets spending limits, from a single task up to a whole department. The stop itself goes on the record, timestamped. It shows who stopped what and why, and what the work was doing at that moment. The control and cost page shows the screen and the limits.

An untested kill switch can't be relied on, which is why stop drills are a published measure, with the date of each drill and the time it took to stop each Virtual Employee and each department. Our piece on the AI agent kill switch sets it against the questions bank examiners now ask.

Blocked attempts are the number to watch

#

Every control above produces refusals, and the question is what happens to them. In many systems a refused action is dropped quietly, and the security team learns of the attempt late, if at all, from a log nobody reads.

Here an out-of-policy attempt is blocked and logged, then shown on the control screen. It is never silently dropped. The count of blocked attempts is published beside the stop drills because it tells a security team two things at once: the controls are working, and something is pushing on them. A count that rises after a new document source goes live deserves a look before anyone asks.

A plan is checked before it runs; an out-of-policy step is blocked, logged and shown on the control screen, and added to the blocked-attempt count
Each blocked step is logged and shown.

What the auditor sees

#

Each task leaves a receipt. It names the identity that did it and the checked plan; the model and the rules it ran under; what it touched; and the cost and the outcome. Phylax keeps every receipt and approval timestamped and tamper-evident, along with each queue move, and sessions on your network are recorded. When an examiner asks who approved a step, or whether anything has changed since, the answer is already in the record. Our guide to reading an AI task record walks through one line by line, and lenders will find the same controls set against the examiners' questions in the CSBS AI framework checklist.

Five questions to put to any vendor

#

Two public references are worth having open as you do: the OWASP Top 10 for Agentic Applications (OWASP GenAI Security Project, December 9, 2025), which counts identity and privilege abuse among the main risks, and the NIST NCCoE concept paper on software and AI agent identity and authorization (NIST, February 5, 2026).

Before you sign an access request for any AI agent, ours included, ask:

  1. Is each agent its own identity in our directory, or a shared service account?
  2. Where are its credentials held, and how long do they live?
  3. What checks a plan before it runs, and can a document it reads change those checks?
  4. How fast can we stop one agent, and how fast can we stop all of them?
  5. When a check blocks an action, do we see it, or is it dropped?

Our answers to all five are on the security and trust page, and the wider questions buyers ask are in our Virtual Employees FAQ. Bring your security questionnaire to a briefing, and we will walk your team through each control.

Share

Questions

What people ask.

Is it safe to give an AI agent a login?

It is, when the login behaves like an employee's. Each Virtual Employee gets its own identity in your directory with least privilege, its credentials are short-lived and held by a gateway, every plan is checked before it runs, and you can stop it at once. Disable the account and the work stops, and every action is tied to one identity and one checked plan.

What is AI agent identity and access management?

It is giving each AI agent its own identity, permissions and credentials, and governing them the way you govern a person's. For a Virtual Employee on your network, that means an account federated to your identity provider, so your conditional access rules, privileged-access rules and access reviews apply to it, and every action it takes is tied to one identity and one checked plan.

How do we stop an AI agent that has a login?

From one screen, at once. With Inovagio, an operator can pause or stop one Virtual Employee, a department or everything at once, and the stop is recorded with who did it, when, why and what the work was doing. Stop drills, with the time each one took, are a published measure, so the switch is tested rather than assumed.

Can a malicious document make an AI agent do something it shouldn't?

It can change what the Virtual Employee proposes, not what the checks allow. Every plan is checked against the Virtual Employee's role, your approved rules and the data it has been granted before anything runs, and the checked plan runs inside a sandbox. An attempt outside policy is blocked, logged and shown on the control screen, and it adds to the published count.

Where are an AI agent's passwords and credentials stored?

Nowhere the AI can read them. The gateway holds short-lived credentials, scoped to one step, issued when the step starts and revoked when it ends. They never appear in a prompt and are never written to disk, and every request that passes through the gateway is logged against the Virtual Employee that made it.

The products, industries and examples it covers.

See it on your work

See how one of your tasks runs, start to finish.

A 30-minute walkthrough of one of your workflows, the approval step and the audit record, with the people who build the platform.

We use these details to answer your request. See our privacy policy.

What happens next

  1. 1We reply within one business day to find a time.
  2. 2A 30-minute walkthrough of one of your tasks, start to finish.
  3. 3You leave with a short write-up of what we covered.