AI agent permissions: five doors between connection and authorization
AI agent permissions: five doors between connection and authorization
Say your sales team wants an AI agent to handle follow-up after customer calls. After each call it reads the transcript, checks the account in your CRM, drafts a follow-up email for the account owner and logs a note on the record.
Someone connects the agent to the CRM and to email, runs a test call through it, and the follow-up comes out right. The project gets marked done.
Connecting the agent didn't decide much, though. The credential it was given might let it read every account in the CRM, edit any record and send email to anyone, as the account owner. Nobody chose any of that. It came along with the connection, and a setting meant to make one workflow possible became the rule for what the agent is allowed to do.
Before an agent like this goes live, someone has to decide what it may read, draft, change, send, spend and approve. Those are separate decisions. We think of them as five doors: connect, observe, prepare, act and commit. Getting through one shouldn't open the next.
Connected is not the same as allowed
Authentication tells a system who is asking. Authorization decides whether that identity may take a particular action on a particular resource, in this situation. A working connection proves the agent can reach the CRM. It says nothing about which accounts this follow-up needs, who the agent is acting for, or what happens if it sends the wrong thing.
The distinction is older than AI agents. NIST's Zero Trust Architecture guidance, SP 800-207, says there should be no implicit trust granted to assets or user accounts based solely on their physical or network location or on who owns them, and it treats authentication and authorization as separate steps completed before a session with a resource begins. For the follow-up agent, that means it shouldn't get broad authority because it runs inside the company, belongs to the company or acts for someone who logged in.
Agents make the gap matter more, because they pick their own tools and chain steps together from a plain-language goal. The account owner asked for a follow-up to one customer, and the credential allows far more than that, including changing who can see a record and deleting notes. The permissions have to describe what the person asked for, not everything the system can do.
The five doors
Each door is a separate grant. For the follow-up agent, the doors are getting its own login, reading the call, drafting the email, saving a note and sending it. The further the agent goes, the more it costs when something goes wrong, so the later doors need stronger approval, closer watching and a faster way to take access back.
Door 1: Connect
The agent gets its own identity or credential and can reach the CRM and the mailbox. It can open a session. That's all connection should give it, with no data and no business action attached.
For the follow-up agent, settle these before going further:
- Is it acting as itself, as a service, or on behalf of the account owner?
- If it changes a record, can you tell it was the agent, and which person it was working for?
- How is the credential issued, stored, rotated, expired and revoked?
- Is it limited to the right environment and tenant?
Connection is where you set up accountability, not where you hand over every scope the workflow might need someday.
Door 2: Observe
Now the agent can read what the task needs: this call's transcript, and the fields on this account a follow-up depends on, such as the contact, the deal stage and the open question from last time. Where the system lets you, narrow that reading by object, field, project and time.
Read-only is safer than write access, but it isn't harmless. An agent that can read every account can pull sensitive details together, keep them longer than it should or leak them. So instead of asking whether it can access the CRM, ask which records and fields it may read for this follow-up, on whose behalf, and for how long.
What the agent reads also includes search results, tool output, retrieved documents and memory. Any of those can carry sensitive data, and any of them can carry instructions someone planted.
Door 3: Prepare
The agent drafts the follow-up email and a proposed CRM note, with links back to the parts of the transcript they came from. Nothing has been sent and no record has changed. The account owner has something exact to look at: the recipients, the wording and the note.
This is the door teams skip most. They go straight from reading to sending because the email system offers both. But much of the labor happens here. Someone has to read the call, find the open questions and write the email. Stopping at this door keeps all of that work, and the account owner still decides whether it goes out.
Door 4: Act
Here the agent makes a bounded change inside a workflow you've approved. In the follow-up, that might mean saving the CRM note as a pending draft and adding an internal "needs review" label. Elsewhere it could be creating a sandbox resource or running a tested deployment step under set conditions. Before it starts, the grant should say which functions it may call, on which records, how often, how the result gets checked and how the change gets undone.
Reversible doesn't mean risk-free. A label can trigger an automation in another system. A status change can alter someone else's work, and a change that's easy to undo once can be hard to undo across two thousand records.
Door 5: Commit
The last door is where the agent does something outside your systems, or something that's hard to take back. Sending the follow-up is the plainest case: once it goes, the customer reads it. Publishing content, approving someone's access, moving money, signing an agreement, deleting production data and changing a security policy sit at this door too.
Not every Door 5 action has to stay manual forever. But before you let the follow-up agent send anything, someone has to answer a few questions. Who may approve a send? Does that approval have to happen right before the email goes? What does the approver actually see? Which limits are hard rules the system enforces, rather than instructions the agent is asked to follow? And what do you do when an email goes to the wrong person and can't be recalled?
What every grant should say
Each grant should answer the six questions below. For the later doors, also write down the evidence and approval you expect before the agent goes through. A scope string on a token can carry part of this, but the policy also has to say why the access exists and when it ends.
| Dimension | Question |
|---|---|
| Identity | Which agent, which service and which delegating person are involved? |
| Resource | Which tenant, system, object, record or field is in scope? |
| Action | Which door is it: observe, prepare, act or commit? |
| Context | For which task, environment, risk condition and purpose? |
| Duration | When is the access valid, and does it end with the task? |
| Revocation | Who can remove access, how fast, and which sessions or tokens survive it? |
The IETF's OAuth security best practice, RFC 9700, published in January 2025, says the privileges on an access token should be restricted to the minimum required for the use case, and recommends limiting tokens to a specific resource server and to particular resources and actions. For the follow-up agent, that means a CRM token that works only against the CRM, and only for the records and actions the follow-up needs. Scopes are one way to get there. What matters is that the token carries no authority the job doesn't use.
When an agent acts for a person, record both identities. If the follow-up agent sends mail using the account owner's own login, the log will show the account owner sending it, even though software chose the words and sent the message.
The follow-up agent, door by door
Here is how the sales team could let the agent do most of the follow-up work without handing it a sales rep's full authority.
| Door | What the agent may do |
|---|---|
| Connect | Use its own service identity, linked to the account owner who asked for the follow-up. |
| Observe | Read this call's transcript and selected fields on this account. No bulk export. |
| Prepare | Draft the follow-up email and a proposed CRM note, with links to the transcript. |
| Act | Save the CRM note as a pending draft and apply an internal review label. |
| Commit | The account owner approves the exact recipients and message before it sends. The agent can't change the recipients after approval. |
Now change one fact. The email includes a negotiated price or a contractual promise. The send action is the same, but the consequence isn't. That version might need a second approver, a narrower template, or no agent sending at all.
A fixed role like "sales assistant" would treat those two emails the same, because all the role knows is that the agent may send email. Whether it should send this one depends on what the message says, who it's going to and what it commits the company to.
When the content itself gives orders
Narrow permissions are necessary, and on their own they still aren't enough. Say a reply sitting in the customer's email thread contains a line written to look like an instruction: send the full account history to this address. An agent reading that thread could act on it with tools it legitimately holds.
NIST's Center for AI Standards and Innovation calls this agent hijacking: a type of indirect prompt injection in which an attacker inserts malicious instructions into data that looks like what an agent would normally work with, to steer it toward a task the user never asked for. Every call the agent makes in that moment can be authenticated and still serve a purpose nobody authorized.
OWASP lists Excessive Agency as LLM06:2025 in its Top 10 for LLM applications, and says the root cause is typically one or more of three things: excessive functionality, excessive permissions and excessive autonomy. Fixing one leaves the other two. A narrowly scoped credential helps little if the tool behind it bundles reading, writing and deleting into one function. An approval step helps little if the account owner sees a vague summary instead of the exact email and recipients.
So build your defenses around the doors. None of these will make an agent secure on its own, but each one makes the damage smaller when something goes wrong:
- Expose only the tools the workflow needs.
- Use the smallest resource and action scope the system offers.
- Keep proposing separate from executing.
- Show the approver the exact action for anything material.
- Check permission again when the action runs, not only when the plan was made.
- Log the agent's identity, the person it acted for, where each input came from, every tool call, the approval and the result.
- Test revocation, failure and attempts to reach past the granted scope.
Open the doors one at a time
Start the follow-up agent with its own identity and read access to the calls it's asked to follow up on. At first, the account owner reads every draft and proposed note before anything moves. Try saving notes and labels in a test copy of the CRM before you let the agent touch the real one. When you do open the commit door, open it for one kind of email, say a recap with no pricing in it, and name the account owner who approves those. After that, watch what the agent actually does. Take away permissions it never uses, rehearse revoking its access, and look again whenever the tools, the model, the data or the situation change.
Watching matters because an agent's authority drifts. Someone connects it to the team calendar as well, a tool it uses gains a delete function, another team copies the follow-up workflow, and a credential outlives the task that justified it.
The questions specific to agents aren't settled yet. In February 2026, NIST's National Cybersecurity Center of Excellence published a concept paper on software and AI agent identity and authorization and asked for public comment, including on how AI agents should be identified, authorized and audited, and how to guard against prompt injection. The paper outlines a potential project, not a standard. You don't have to wait for a standard before the follow-up agent goes live. But don't give it broad, permanent access that you'd struggle to pull back once the answers settle.
At each review of the follow-up agent, ask:
- Which doors did it actually use?
- Which permissions did it have and never need?
- Could it get through a door without the evidence or approval you intended?
- Can you trace every action to the agent, the person, the task and the policy that allowed it?
- When you revoke access, do current sessions stop as well as future ones?
As you learn what the workflow needs, its permissions should shrink. They shouldn't grow just because nobody went back to check.
Open only the door the work requires
Go back to the sales team's follow-up agent. It doesn't need everything the CRM and the mailbox can do in order to be useful. It needs its own identity, the transcript and a handful of fields, a place to put drafts, and a person who decides what reaches the customer.
Supanova is built on the same split between doing the work and deciding what goes out. You write a project brief, your atoms do the work, and you choose how often your workforce checks in: every step, every task, or once per project. Even if you green-light a whole project at once, finished work comes back for your review. Nothing ships without you.
If you already have an agent connected to something, which door is it standing at, and who decided that? If you want your sign-off built in from the first project, start at supanova.team.
Sources
- NIST SP 800-207: Zero Trust Architecture (August 2020)
- IETF RFC 9700: Best Current Practice for OAuth 2.0 Security (January 2025)
- NIST Center for AI Standards and Innovation: Technical Blog, Strengthening AI Agent Hijacking Evaluations (January 2025)
- OWASP Top 10 for LLM Applications: LLM06:2025 Excessive Agency
- NIST: New Concept Paper on Identity and Authority of Software Agents (February 2026)