What AI agent authorization is, and why user permissions are not enough
✓
The four questions every agent permission check must answer
✓
Seven design rules for agent identity, scopes, tenants and approvals
✓
When to use RBAC, ABAC or relationship-based access control (ReBAC) for agents
✓
How to sort tools into risk tiers and attach approval rules
✓
The permission mistakes we see most often in SaaS agents, and how to fix them
The moment an AI agent can act inside your product, it needs an answer to a question your permission system was never designed to ask: what is this agent allowed to do, for this user, in this customer's account, right now?
AI agent authorization is the set of rules and checks that decide what an agent may do, on whose behalf, with which data and in which tenant, at the moment it acts. It sits on top of your existing user permissions, but it is not the same thing. An agent chains actions, reads untrusted content and can be talked into things a user would never do.
Security leaders have noticed. In our own search data, the questions buyers type into AI assistants increasingly read like this: agents now hold permissions, reach sensitive systems and take real actions, so how do we keep them least-privileged and visible? This guide answers that for B2B SaaS teams. It covers the design rules, the access models and the mistakes we fix most often, and it builds on our complete guide to AI agents for SaaS.
Why AI agent permissions are different from user permissions
In a classic SaaS product, a permission check is simple: this user, this role, this resource. Agents break that model in three ways.
Agents act on someone's behalf. When an agent updates a record, the system needs to know both who the agent is and which user it is working for. Without that, every action looks like it came from a generic service account.
Agents chain actions. One request can trigger ten tool calls across several systems. A permission that is harmless for one call can be dangerous in combination.
Agents read untrusted content. An email, a PDF or a web page can contain instructions. If the agent has broad access, a hidden instruction can turn it into what security engineers call a confused deputy: a trusted program tricked into misusing its own authority.
That is why the OWASP Top 10 for Agentic Applications, published in December 2025, puts tool misuse (ASI02) and identity and privilege abuse (ASI03) near the very top of its list.
Security research
2 of 3
of the top risks in the OWASP Top 10 for Agentic Applications are about what agents can access: tool misuse, and identity and privilege abuse.
The four questions every agent permission check must answer
Every time an agent calls a tool, your system should be able to answer four questions before the call runs.
Who is acting? The agent's own identity, plus the user it is acting for.
In which tenant? The customer account the action belongs to, taken from the authenticated session, never from the prompt.
What action on which resource? Read this invoice, update that contact, send this email.
Under which conditions? Limits, time windows, data sensitivity, and whether a person has to approve first.
If any of the four answers is missing, the safe default is to deny the call and log why.
7 rules for designing AI agent permissions
1. Give every agent its own identity
Treat each agent as a non-human identity with its own credentials, owner and lifecycle. Never let agents share an admin API key. A distinct identity is what makes every later control possible: scoping, monitoring, revocation and an audit trail that says which agent did what.
2. Act on behalf of a user, with the intersection of rights
When an agent works for a user, its effective permission should be the intersection of what that user may do and what the agent is designed to do. A finance manager may approve payments; an invoice-matching agent working for her should still only be able to read invoices and propose matches.
3. Use short-lived, narrowly scoped tokens
Issue tokens per task or per session, scoped to the tools and the tenant involved, and let them expire in minutes, not months. Standard OAuth flows support this: token exchange (RFC 8693) lets a service swap a user's token for a narrower one the agent can use. The Model Context Protocol builds its authorization on OAuth 2.1 and treats each MCP server as a protected resource, so agents and tools can follow the same pattern.
4. Enforce tenant isolation inside the tool
The tenant must come from the authenticated context, never from text the model produced. Every tool call is filtered by that tenant at the data layer. If a prompt says "look up the invoices for customer B" while the session belongs to customer A, the tool simply cannot see customer B.
5. Choose the right access model for resources
Model
How it decides
Good for agents when
Limits
RBAC (role-based)
Permissions follow the user's role
Your product already has clear roles and the agent mirrors one
Too coarse for record-level rules
ABAC (attribute-based)
Rules on attributes: amount, region, sensitivity, time
You need conditions such as "payments under 5,000" or "EU data only"
Rules can get hard to reason about
ReBAC (relationship-based)
Access follows relationships: owner of, member of, assigned to
Access depends on who works with which account, project or patient
Needs a dedicated authorization service
Most SaaS products end up combining them: roles for the broad strokes, relationships for record-level access, and attributes for the conditions that matter to agents. Relationship-based systems such as Google's Zanzibar design and its open-source descendant OpenFGA are a good fit when agents need to respect fine-grained sharing.
6. Sort tools into risk tiers
Not every tool deserves the same treatment. We classify each one before it is exposed to an agent.
Tier
Examples
Default rule
Read
Search records, view an invoice, read a document
Allowed within scope, logged
Write
Update a field, create a draft, tag a ticket
Allowed within limits, reversible, logged
External
Send an email, post a message, call a partner API
Approval or a pre-approved template
Destructive or financial
Delete data, approve a payment, change permissions
Always a human approval, with the evidence shown
7. Log every decision, allowed or denied
Record the agent identity, the user, the tenant, the tool, the parameters, the decision and the reason. Denied calls matter as much as allowed ones: a spike in denials is often the first sign of a prompt injection attempt or a badly designed tool.
Where to enforce permissions: in code, not in the prompt
A system prompt is not a permission system. Instructions such as "never access other customers' data" help the model behave, but they can be overridden by clever input. Enforcement has to live in code the model cannot rewrite: the API gateway, the tool implementation and the data layer.
In practice that means three layers. The gateway authenticates the agent and the user and attaches the tenant. The tool checks the action against the authorization service. The data layer filters every query by tenant and scope. The model only ever sees the results it was allowed to see.
The AI agent permission mistakes we fix most often
Mistake
Why it hurts
Fix
One admin key for all agents
Any compromise exposes every tenant
Per-agent identities and scoped, short-lived tokens
Tenant ID passed in the prompt
Prompt injection can switch tenants
Tenant taken from the authenticated session only
Agents inherit all of the user's rights
An assistant can do everything its user can
Intersection of user rights and agent scope
Permissions described only in the system prompt
The model can be talked out of them
Enforcement in gateway, tools and data layer
No log of denied calls
Attacks and design gaps stay invisible
Log every decision with a reason
How we design agent permissions at Linnify
In our ARC (Agentic Release Control) framework, permissions are designed before the agent is built. In the Ingest phase we write down each agent's tools, decision boundaries and approval points with the product owner and the domain expert. In the Deploy phase we implement access control, role management and logging as part of the release, not as a patch afterwards. Because the agent runs in our client's own infrastructure, their security team can review every control with the tools they already use.
From our work: multi-tenant accounting
Agents that respect each customer's workspace
A Munich-based cloud ERP gives each of its 400+ property management customers a dedicated workspace with its own financial logic. The accounts payable and bank reconciliation agents we built had to work inside that structure, for every tenant, without crossing between them.
Give each agent its own identity, let it act only within the overlap of its own scope and its user's rights, keep tenants apart in code, and put a person in front of anything external, destructive or financial. Agents built that way can be given more autonomy over time, because every action they take is visible and bounded.
Agent security
Adding agents to a multi-tenant SaaS product?
Talk to the Linnify team about agent identity, tenant isolation and approval design before your first agent reaches customers.
AI agent authorization is the set of rules and checks that decide what an AI agent may do, on whose behalf, with which data and in which tenant, at the moment it acts. It builds on user permissions but adds the agent's own identity, scope and approval rules.
Should AI agents have their own identity?+
Yes. Each agent should be a distinct non-human identity with its own credentials, owner and lifecycle. Shared admin keys make it impossible to scope, monitor or revoke a single agent, and they hide which agent did what.
How should an AI agent get permissions in a multi-tenant SaaS product?+
The agent should act on behalf of a signed-in user, with an effective permission equal to the overlap of that user's rights and the agent's designed scope. The tenant comes from the authenticated session, and every tool call is filtered by it at the data layer.
What is the confused deputy problem for AI agents?+
A confused deputy is a trusted program that is tricked into misusing its own authority. For agents, it usually happens through prompt injection: hidden instructions in an email or document push an over-privileged agent to read or change data it should not.
Can I restrict an AI agent with its system prompt?+
Only as a second line of defense. Instructions help the model behave, but they can be overridden by malicious input. Permissions must be enforced in code the model cannot change: the gateway, the tools and the data layer.
How does MCP handle authorization?+
The Model Context Protocol builds its authorization on OAuth 2.1 and treats each remote MCP server as a protected resource. Hosts obtain scoped tokens for each server, so tool access can follow the same least-privilege rules as the rest of your product.
References and further reading
• OWASP GenAI Security Project. (2025, December). OWASP Top 10 for Agentic Applications, reported by TechRepublic. techrepublic.com