Picture the controller at a 60-person firm who sets up an AI agent to chase vendor invoices. She connects it to her mailbox, her accounting login and the shared drive, and it works. It reads every message she can read, opens every folder she can open and posts entries under her name. Nobody created an account for it, nobody wrote down what it’s allowed to do, and the audit log can’t tell her actions from the agent’s. That scene is playing out across the mid-market right now, and the rules for AI agents under NIST SP 800-171 already answer it. The standard just calls them something else.

The standard already treats an agent as a user

NIST SP 800-53 Revision 5 defines a user as an “individual, or (system) process acting on behalf of an individual, authorized to access a system.” NIST SP 800-171 Revision 2, the version the Department of War enforces for defense suppliers, builds the same idea into its first requirement. Requirement 3.1.1 limits system access to “authorized users, processes acting on behalf of authorized users, and devices.” Requirement 3.5.1 says to “identify system users, processes acting on behalf of users, and devices,” and 3.5.2 says to authenticate them before they get in.

An AI agent that reads a mailbox, queries an ERP system or writes to a file share is a process acting on behalf of a user. The requirement was written before these tools existed, and it covers them anyway. What’s new is that the process now decides for itself which actions to take, which makes the identity, the permissions and the log matter more, not less. For any regulated company, the 110 requirements in Revision 2 are the cleanest short checklist I know of for software that acts on its own.

The shared login is where it breaks

Most agents get set up on a person’s credentials because it’s fast. Measured against the standard, that one shortcut fails four requirements at once. Requirement 3.3.2 says the actions of individual users must be “uniquely traced to those users,” and when the agent and the controller share a login, nothing in the log belongs clearly to either of them. Requirement 3.1.5 calls for least privilege, and an inherited login carries every permission that person accumulated over the years. Requirement 3.5.3 requires multifactor authentication for network access to accounts, and agents usually sidestep it with a long-lived token or an app password that never expires. Requirement 3.5.6 says to disable identifiers after a period of inactivity, which nobody applies to a token sitting in a config file.

The data points the same way. IBM’s 2026 Cost of a Data Breach Report found that 92 percent of organizations that suffered an AI-related breach lacked proper access controls, and that security incidents involving shadow AI more than doubled in a year, to 43 percent from 20 percent. Verizon’s 2026 Data Breach Investigations Report found that 45 percent of employees are now regular users of AI on corporate devices, up from 15 percent, and it singles out service and machine accounts for special attention because those are the accounts an agentic future will run on.

Six rules, applied to an agent

Here is how my team at Stealth Technology Group applies the standard to an agent, with the requirement each rule maps to.

  1. Give it an identity (3.5.1). Every agent gets its own account, a name that says what it is, and a named human owner. Nothing runs as a person.
  2. Give it a credential of its own (3.5.2, 3.5.10). A scoped token or certificate issued to the agent, kept in a vault, rotated on a schedule and revocable in one step. Never the owner’s password.
  3. Scope it to the job (3.1.2, 3.1.5). An invoice agent reads the invoices mailbox and posts to accounts payable. It doesn’t read the CEO’s inbox or approve payments.
  4. Keep it out of privileged functions (3.1.7). An agent never holds administrator rights. If a task needs a privileged action, a person performs it and the log captures it.
  5. Log everything it does under its own name (3.3.1, 3.3.2). The record shows the agent’s identity, the owner it acts for, the data it touched and the action it took.
  6. Retire it on purpose (3.5.5, 3.5.6). Disable the identity when it’s idle, remove it when the owner leaves, and never reuse its identifier for something else.

None of this needs a new platform. Your identity provider, permissions and logging already do this for people. The work is deciding to do it for software.

If you hold controlled data

For a defense supplier, an agent that touches controlled unclassified information is inside the assessment boundary. The Department of War suspended CMMC Phase II on July 13, 2026, but says it will enforce NIST SP 800-171 Revision 2 through self-assessments and select government-led assessments during its review, and that the suspension does not eliminate the requirement to protect information under DFARS 252.204-7012. That clause still requires reporting a cyber incident within 72 hours of discovery, and an agent that sends a controlled drawing somewhere it shouldn’t is an incident. The annual affirmation, which I covered in what the CMMC executive affirmation attests to, covers the agents too.

What to do this quarter

Start with the inventory, the same five-day pass I laid out in What is shadow AI?, and add one column: does this tool act on its own? Every yes gets an owner, an identity, a permission scope and a line in the log within 30 days. Put the count and the exceptions on the one-page AI brief your board reads each quarter.

Frequently asked questions

Are AI agents users under NIST SP 800-171?

Yes. Requirements 3.1.1 and 3.5.1 of Revision 2 cover “processes acting on behalf of users.” An agent that accesses your systems is covered the same way an employee is.

Should an AI agent have its own account?

Yes. An agent needs its own identity, its own credential and a named owner, with permissions limited to its task. Running it under a person’s login breaks the traceability requirement in 3.3.2 and the least-privilege requirement in 3.1.5.

Does the CMMC pause change the rules for AI agents?

No. The Department of War says it enforces NIST SP 800-171 Revision 2 through self-assessments during its review, and DFARS 252.204-7012 still applies, including 72-hour incident reporting.

My team at Stealth Technology Group builds this control set into our AI strategy and governance work for regulated mid-market companies and defense suppliers. If you’d like to talk through the agents already running in your company, contact me here.

3 Responses