Agentic AI
AI Agent Permissions: A Plain Guide to Setting Them Safely.
AI agentNoun · glossaryAn AI agent is a focused worker you hand a single job to.Read the full entry → permissions decide what an agent may see and what it may do. Give each agent its own identity. Give it the least access it needs. Make a human approve anything risky or hard to undo. Then enforce those rules outside the AI model, and keep logs.
Treat an agent like a new hire. You would not hand them every key on day one. You give them the keys for the job. Agents need the same rule.
TL;DR
- Permissions are the rules for what an agent can reach and what it can do.
- Give each agent its own identity. Never let it borrow a person's login.
- Start with read-only access. Add more only when the job needs it.
- Humans approve anything destructive, external or hard to reverse.
- Enforce the rules in a control layer outside the model, and log every action.
What AI agent permissions are
A permission is a yes or no. Can the agent read this file? Can it send this email? Can it change this record?
One security write-up describes agent permissions as the authorisations that decide which resources an agent can reach and which operations it can perform. The real scope comes from the logins, keys, tools and services the agent touches.
This part trips people up. An agent can look limited. But if its tool has a master key behind it, the agent has a master key.
Why agents need their own rules
A normal app does what its code says. An agent picks its own steps. It can chain tools together in ways nobody planned.
So agents need a permission model of their own. They are hard to predict. A hidden instruction in a web page or file can trick them. And they string tool calls together in new ways.
Think of a kitchen. A normal app is a cook following a recipe. An agent is a cook who can open any cupboard. Put locks on the cupboards that matter.
The risks you are guarding against
You do not need to learn every threat. Know the main ones. These are the ones tied to permissions, as OWASP lists them:
- Prompt injection. Hidden text tells the agent to do something else.
- Tool abuse and privilege escalation. The agent uses a tool in a way nobody meant, and gains more power.
- Data leaks. Private data goes where it should not.
- Excessive autonomy. The agent acts too freely.
- High-impact actions. Deleting, paying or sending without a check.
Microsoft names five problems in its least privilege guidance for agents:
- Unclear identity.
- Permission creep (access that slowly piles up).
- Tools with too much reach.
- Weak audit trails.
- Slow shutdown when something goes wrong.
Most of these come down to access that is too wide. That part you can fix.
Why this is not a small problem
The Cloud Security Alliance reports that 82% of organisations found AI agents they did not know about. It also says 65% had an agent security incident in the past twelve months. And only 21% have a formal process to retire agents.
So the first job is simple: know which agents you have.
On 30 April 2026, CISA and four allied agencies (Australia, Canada, New Zealand and the UK) released joint guidance on Agentic AINoun · glossaryAgentic AI describes systems that pursue a goal over multiple steps rather than answering one prompt at a time.Read the full entry →. It is the first joint guidance of its kind on the risks of autonomous agents. It groups the risks into privilege, design and configuration, behaviour, structure and accountability.
Step 1: Give every agent its own identity
An identity is the agent's name badge. It shows who is acting.
Do not run an agent on your own login. Then every action looks like you did it. You cannot tell the agent's work from yours.
AWS advises distinct service identities, short-lived credentials and permission boundaries. The aim is the minimum permissions each task needs.
Short-lived credentials are like a hotel key card. It works for a short time, then it dies. A stolen card is not much use a day later.
Avoid these:
- Shared accounts used by many agents.
- Static keys pasted into code and never changed.
- One "admin" login for everything.
The CSA's reading of the CISA guidance is blunt: static API keys and shared service accounts do not fit the guidance. Each agent should carry a unique identity with short-lived credentials.
One identity is not enough
An identity tells you which agent acted. It does not tell you a person approved the action. One identity-security writer says that second problem is the one with legal and financial weight.
So name an owner for every agent. A real person, who answers for what the agent does. The Five Eyes guidance says the same: define who is responsible and keep humans accountable. If you want a wider view of that duty, read about agentic AI risk management across a whole organisation.
Step 2: Give the least access that works
Least privilege means the agent gets only what the task needs. Nothing extra.
NIST published its design tenets for agent identity on 27 October 2026. It says agents should hold no standing rights when idle. This is called zero standing privilege. Access should be limited to the minimum scope for one task. And one agent should not hold conflicting powers.
In plain words, the agent asks for a key when it needs one. It gets the key for that job. Then the key goes away.
NIST also says access should be just-in-time, time-limited and tied to one audience. Each tool call should be checked as it happens. If something looks odd, the system should be able to cut the agent off.
A simple way to scope access
Ask four questions for each agent:
- Which data? Name the folders, tables or records. Not "everything".
- Which tools? List them. Remove any it does not use.
- Which actions? Read, write, delete, send or pay. Pick the smallest set.
- For how long? Set an end time on access.
Then write the answers down. A short list beats a clever system nobody reads.
Roles, or rules that read the situation?
Role-based access gives permissions by job title. It is easy to grasp. We cover it in our plain guide to role-based access for AI systems.
But roles are fixed. The CSA says rules based on attributes and policies fit zero standing privilege better than fixed roles. They check the task, the data sensitivity and the conditions at the moment of the request.
Here is an example. A role says, "Support agents can read customer files." A context rule says, "This agent can read this one customer's file, while it handles this one ticket." The second is tighter.
Start with roles if you are small. Move to context rules as you grow.
Step 3: Sort actions into approval tiers
Not every action needs a human. If it did, you would not save any time. So sort actions by risk.
One common pattern uses three tiers. It comes from a write-up on approval and autonomy:
| Tier | Type of action | Who decides |
|---|---|---|
| 1 | Read-only (look, search, summarise) | Agent acts alone |
| 2 | Scoped writes (edit a draft, update a field) | Agent acts alone if trusted, else a human approves |
| 3 | Destructive or external (delete, send, pay) | A human always approves |
Singapore's model framework for agentic AI agrees on the top tier. High-stakes or irreversible actions such as editing sensitive data, deleting data, sending messages or making payments should need human approval.
How to place your own actions
Take your agent's task list. Put each action in a tier. Use two tests:
- Can I undo it? If not, move it up a tier.
- Does it leave the building? Emails, posts and payments reach other people. Move them up.
A map of your repeat tasks helps here. This automation opportunity map is one way to list what could be automated before you decide how much freedom each piece gets.
Avoid approval fatigue
If a person clicks "approve" a hundred times a day, they stop reading. Then the check is fake.
Keep approvals rare, and put them only on Tier 3. Njin has a good example of where an agent should stop and hand over, in its post on AI sales agent handover rules. Show the person what the agent wants to do and why it matters. A good prompt says: "Send this refund of $400 to this customer." A bad one says: "Allow action?"
Step 4: Enforce the rules outside the model
Most teams miss this step. Do not rely on the AI to follow its own rules. A model can be tricked. A line in the instructions saying "never delete files" is a request, not a lock.
OWASP's advice is to enforce authorisation in the execution component, outside the agent's context. That component should check that an approval:
- belongs to this actor and this exact tool call,
- is still valid,
- has not been used already.
It should check and use up the approval in one step, just before the action runs. And it should fail closed. That means unknown tools and missing approvals get a "no".
Back in the kitchen: the cook can ask for the knife. But a manager decides if it leaves the drawer, and the cook cannot talk the manager round.
Security writers put it this way: decisions about runtime access belong in infrastructure, not in the agent or the app. The agent proposes. The control layer decides.
We also wrote about why one lock is never enough. Stack the layers: identity, scoped access, approvals, a control layer and logs.
Step 5: Log, watch and be ready to switch off
Permissions are not set and forget. Access drifts, jobs change and old agents stay on.
The Singapore framework suggests you monitor behaviour through detailed logs and set alerts for anything odd. It also says to audit permissions often and remove stale access. And agents need a kill switch or rollback for real-time action.
A simple routine:
- Every action: log who, what, when and on whose behalf.
- Every week: scan for odd spikes, new tools or failed attempts.
- Every quarter: review each agent's access. Cut what it did not use.
- On change: when a person leaves or a job ends, retire the agent.
- Always: know how to switch an agent off in minutes.
Test the kill switch before you need it.
A 30-day starter plan
You can get this done in a month.
Week 1: Find your agents. List every agent, tool and key. Name an owner for each.
Week 2: Fix identities. Give each agent its own login. Swap static keys for short-lived ones.
Week 3: Set tiers. Sort each action into Tier 1, 2 or 3. Add approvals for Tier 3.
Week 4: Add logs and a switch. Turn on logging. Write a one-page shutdown plan. Test it.
If you want a deeper plan for rules and oversight, see our 90-day playbook for agent governance. For written rules your team can follow, read about AI policy development.
Common mistakes
These come up again and again:
- Using a personal login. You lose track of who did what.
- Starting wide. Start narrow. Widen later.
- Trusting the prompt. Instructions are not locks.
- Approving by habit. Too many prompts mean none get read.
- Forgetting old agents. Unused access is a free door.
Okta's guidance lists the same traps: excessive agency, privilege creep and long-lived credentials.
How we run it
We treat AI as an enthusiastic intern, not a magic button. So we run 10/80/10. A person plans for the first 10%. The AI does 80%. A person checks and makes it right for the last 10%. Nothing goes out without a human yes.
That is a permission rule. The AI drafts. A person approves.
We also learned the account lesson the hard way. The moment you run AI across a team, your AI accounts become a target, the same as a bank login. So we lock them down. Individual logins, not shared magic links. 2FA on everything, with an authenticator app over SMS. Hard spend caps, so a breach cannot run up a bill.
The one-shared-login setup most teams run is exactly what a breach feeds on. Fix that first. It costs nothing.
Want help setting this up?
If you run a $1M+ education or consulting business, you may not want to build this alone. We build it with you, as part of the AI Orchestrator OS. Each AI employee gets its own role, its own access and a clear line for when a person steps in.
Prefer to build on your own? See platform access.
Sources
- AI Agent Permissions: Scope, Govern and Enforce Access
- Why AI Agents Need Their Own Permission Model
- AI Agent Security, OWASP Cheat Sheet Series
- Least privilege for AI agents, Microsoft Learn
- CISA Agentic AI Guidance: A Practitioner's Roadmap, Cloud Security Alliance
- Agent identity and permission management, AWS Well-Architected Agentic AI Lens
- Design Tenets for Agent Identity Management Systems, NIST
- The Non-Human Identity Governance whitepaper, Cloud Security Alliance
- Five Eyes Agentic AI Guidance analysis, Cloud Security Alliance
- AI Agent Identity: Why Agents Need Their Own, iProov
- Manual Approval vs Autonomous Tool Execution
- Model AI Governance Framework for Agentic AI, Singapore
- Why AI agents need their own identity, SC Media
- How to implement least privilege for AI agents, Okta
- Automation opportunity map
- Agentic AI risk management
Your next step: list every AI agent in your business today, with one named owner each. Then take the assessment to see where your gaps are.
Frequently asked questions.
James Killick
Founder
The AI Orchestrator. 10+ years building digital products and 200+ apps shipped, now helping $1M+ educators and consultants turn their IP into AI-powered delivery systems.
James Killick founded and runs The AI Orchestrators.
More from James Killick