- Published
- Updated
- 11 min read

TL;DR: When an AI browser agent works inside your accounts, identity and access management has to answer two questions: who is acting, and on whose authority. Give the agent its own identity, scope it to the task, keep phishing-resistant MFA on the person who approves, require approval for high-impact actions, and log every step.
Identity and access management used to have a simple shape. A person signs in, does the work, signs out, and the logs show who opened which door. An AI browser agent breaks that shape. It reads pages, fills in forms and clicks buttons while signed in as someone, often for long stretches, and it takes instructions from whatever text lands in front of it.
That does not make the old identity and access management rules useless. Most of what you need already exists in NIST guidance, CISA's zero trust work and the OWASP lists for AI applications. The job is applying it to a new kind of user. This note walks through how I would set that up for a small or mid-sized organization in Massachusetts, without buying anything new on day one.
What changes for identity and access management when a browser acts for you?

Three identities show up where there used to be one. There is the person. There is the agent, which is software with its own runtime, its own tokens and its own failure modes. And there is the delegation between them: the record that says this agent may do this task for this person, until this time.
Why "signed in as me" is the core risk
The easy way to run an agent is to let it ride on your own browser session. It sees what you see and can do what you can do. That is exactly the problem. OWASP's list of top risks for large language model applications describes indirect prompt injection: it occurs "when an LLM accepts input from external sources, such as websites or files," and that content can alter "the behavior of the model in unintended or unexpected ways." (OWASP GenAI Security Project, LLM01 (opens in a new tab))
Put those together. An agent signed in as your office manager, reading a web page someone else wrote, can be steered into acting with your office manager's access. We covered the attack itself in prompt injection attacks explained. Identity and access management is how you limit the blast radius when one gets through.
Excessive agency is an access problem
OWASP has a name for the other half: excessive agency. It says the "root cause of Excessive Agency is typically one or more of: excessive functionality; excessive permissions; excessive autonomy." (OWASP GenAI Security Project, LLM06 (opens in a new tab)) Two of those three are things your access controls decide. The model's quirks are hard to fix. What it is allowed to touch is not.
What does Massachusetts law expect from access controls?

If your organization holds personal information about Massachusetts residents, 201 CMR 17.00 already sets the floor, and it does not care whether the user is a person or a program. For identity and access management, that is your baseline. Section 17.04 calls for "Secure user authentication protocols," including "restricting access to active users and active user accounts only" and "blocking access to user identification after multiple unsuccessful attempts to gain access." (201 CMR 17.04, via Cornell LII (opens in a new tab))
The same section asks for access measures that "restrict access to records and files containing personal information to those who need such information to perform their job duties" and that "assign unique identifications plus passwords, which are not vendor supplied default passwords, to each person with computer access." It also asks for "Reasonable monitoring of systems, for unauthorized use of or access to personal information." (201 CMR 17.04, via Cornell LII (opens in a new tab))
Read that with an agent in mind. An agent that shares a human login has no unique identification of its own, so the logs cannot separate what the person did from what the agent did. An agent that can reach every file the person can reach is rarely limited to what its job requires. Neither is a new obligation. They are the existing ones, applied to a new kind of account. MyRHC keeps a plain index of Massachusetts compliance sources (opens in a new tab) if you need the wider picture.
How should an AI agent authenticate?

Give the agent its own identity, not your password
NIST's zero trust guidance for cloud-native systems describes the shift plainly: it "requires authentication and authorization policies based on application and service identities in addition to the underlying network parameters and user identities." (NIST SP 800-207A (opens in a new tab)) An agent is a service. Give it a service identity: its own account, its own credentials and its own entry in your identity provider, so it can be turned off without locking out the person it works for. Of all the identity and access management changes agents need, this one does the most.
When the agent acts for someone, the record of that delegation should travel with the request. The OAuth token exchange standard already defines this. Its "act" (actor) claim "provides a means within a JWT to express that delegation has occurred and identify the acting party to whom authority has been delegated." (IETF RFC 8693 (opens in a new tab)) In plain terms, the token says the request is about Dana, and the agent is the one doing it. Your logs then show both names on every action.
Some services now let an agent use a payment or account connection without ever seeing the password. We looked at one example in your AI agent can now log into your Stripe without seeing the password. Prefer that pattern wherever a vendor offers it.
Phishing-resistant MFA for the person who approves
The agent does not need a fingerprint. The human who grants it access does need strong authentication, because that person is the one an attacker will phish. NIST SP 800-63-4, the current digital identity guidelines, says "AAL2 provides high confidence that the claimant controls one or more authenticators that are bound to the subscriber account," that verifiers "SHALL offer at least one phishing-resistant authentication option at AAL2," and that "AAL3 authentication requires a phishing-resistant authenticator." (NIST SP 800-63B-4 (opens in a new tab))
For a small team that means hardware security keys or platform passkeys on the admin accounts that can grant agent access, at a minimum. Text-message codes on those accounts are the weak link.
Session limits still apply

Agents tend to run long, and long sessions are where stolen tokens do the most damage. The NIST guidance sets reauthentication timeouts: a definite overall timeout "SHOULD be no more than 24 hours at AAL2," and "At AAL3, the overall timeout for reauthentication SHALL be no more than 12 hours." (NIST SP 800-63B-4 (opens in a new tab)) Those numbers were written for people, but they are a sensible ceiling for any delegated agent session. If a task needs more than a day, re-approve it.
How do you apply least privilege to an AI browser?

Least privilege is the oldest rule in identity and access management. It matters more for agents, because they act fast, at volume, and never get suspicious of an odd request.
Scope the task, not the person
OWASP's guidance on excessive agency is specific. It recommends you "Limit the permissions that LLM extensions are granted to other systems to the minimum necessary," and that you "Track user authorization and security scope to ensure actions taken on behalf of a user are executed on downstream systems in the context of that specific user, and with the minimum privileges necessary." (OWASP GenAI Security Project, LLM06 (opens in a new tab))
In practice: an agent that books shipping does not need the accounting system. An agent that reads the CRM to draft follow-ups needs read access to contacts, not export rights. Write the task first, then list the systems and actions it needs, then grant exactly that. Everything else is denied by default.
Human approval for high-impact actions

OWASP's advice is to "require a human to approve high-impact actions before they are taken." (OWASP GenAI Security Project, LLM06 (opens in a new tab)) Vendors are building the same idea into their own agent browsers. OpenAI's help page for ChatGPT agent says it "incorporates multiple safeguards, including user confirmations for high-impact actions, refusal patterns for disallowed tasks, prompt injection monitoring," and a "watch mode" that requires user supervision on certain sites. It also advises users to "Avoid typing passwords or private info directly in messages; use takeover mode for sensitive inputs." (OpenAI Help Center (opens in a new tab))
Do not rely only on the vendor's switch. Decide which actions in your business count as high impact (paying, sending to a customer, deleting, changing access) and enforce approval in the system that performs the action, where the agent cannot talk its way past it.
Short-lived credentials kept in a vault
In identity and access management terms, long-lived API keys pasted into an agent's settings are the service account version of a password on a sticky note. Store agent credentials in a secrets manager, issue them for the shortest period the task allows, rotate them on a schedule, and revoke them the day the agent is retired. Treat retiring an agent like an employee leaving. The employee offboarding checklist for IT access (opens in a new tab) I wrote on my own site works almost line for line for an agent account.
Where does zero trust fit in identity and access management for agents?

Zero trust is the frame that makes the rest of identity and access management hang together. NIST describes it as moving defenses "from static, network-based perimeters to focus on users, assets, and resources," and says it "assumes there is no implicit trust granted to assets or user accounts based solely on their physical or network location." (NIST SP 800-207 (opens in a new tab)) An agent running inside your office network gets no extra trust for being there.
CISA's Zero Trust Maturity Model gives a way to measure progress. It "includes five pillars and three cross-cutting capabilities," and within each pillar it gives "examples of traditional, initial, advanced, and optimal zero trust architectures." (CISA, Zero Trust Maturity Model (opens in a new tab)) Identity is one of those pillars. A useful exercise is to score your agent accounts against it separately from your human accounts. Most teams find the agents sit a stage behind. Our knowledge base has a longer explainer on zero trust architecture.
How do you keep a trail of who did what?

Accountability is the part of identity and access management that matters most after something goes wrong. When an agent sends the wrong file to the wrong customer, you need to answer three questions fast: which agent, acting for whom, approved by whom.
That only works if each piece above is in place. A separate agent identity puts a distinct name in the logs. Delegation in the token links that name to a person. Approval records show who said yes. Then send those logs somewhere the agent cannot edit, and watch them. Watching is a topic of its own, covered in our companion note on network security monitoring for AI browser anomalies. Our agentic security article covers the wider set of controls.
What should a small team do first?

You do not need a new platform to start. Here is the order I would work in.
- List every agent. Browser agents, assistants with tool access, automations that sign in to anything. Note which human account each one uses today. Many teams find agents nobody remembered turning on, as we wrote in 82% of companies just found an AI agent nobody remembered.
- Move each agent off shared human logins. Give it its own account in your identity provider, named so a person can recognize it.
- Write down each agent's task and grant only that. Read before write. No admin rights unless the task truly is admin work, and then only with approval per action.
- Put phishing-resistant MFA on every account that can grant agent access.
- Define high-impact actions and require human approval for them in the system that performs them.
- Move agent secrets into a vault with expiry and rotation.
- Send agent activity to central logs and review them on a schedule, along with each agent's access every quarter.
Amelia, my wife and Kief Studio's co-founder and CEO, has written about the governance side, the order to approve things in rather than approving each click, in AI agent governance as a sequence (opens in a new tab). That pairs well with the identity and access management steps above.
Frequently asked questions
Should an AI browser agent use my login?
No, not for anything beyond a quick test. Give the agent its own account so it can be limited, logged and switched off without touching yours, and record which person each task is for. Good identity and access management starts with that separation.
Does an AI agent need multi-factor authentication?
The agent authenticates with credentials or tokens held in a vault. The people who can grant or change its access need phishing-resistant MFA, because they are the accounts attackers will target.
What is the biggest identity and access management mistake with AI agents?
Giving an agent everything its human can reach. OWASP lists excessive permissions as a root cause of excessive agency. Scope each agent to one task.
Does 201 CMR 17.00 apply to AI agents?
The regulation covers access to personal information of Massachusetts residents, with no exception for software. Unique IDs, need-based access and monitoring apply to agent accounts too. Ask your attorney how it applies to you.
How long should an agent's session last?
As short as the task allows. NIST's ceiling for people is 24 hours at AAL2 and 12 hours at AAL3, which is a sensible upper bound for delegated agent sessions.
Which actions should always need a human?
Anything that moves money, sends to a customer, deletes data or changes who has access. Enforce approval in the system that performs the action, not only in the agent.
This article is general information, not legal advice. Kief Studio is not a law firm, and reading it does not create an attorney-client relationship. Laws and regulations change, so check the current text and talk with a licensed Massachusetts attorney about your situation.
Where we fit

Kief Studio sets up identity and access management for organizations that are starting to let AI agents work in their systems. A security architecture review maps every agent and the access it really has. Our zero trust work gives each agent its own identity, scopes it to its task, puts approvals in front of high-impact actions and moves its secrets into a vault. On what we host, managed security watches the logs, and security is included around the clock, as described on LTFI's Watch layer (opens in a new tab).
We work with organizations across Massachusetts, from small offices to regulated teams, and we will tell you plainly where your setup is fine as it is.
First conversation is free. No commitment.




