Skip to contentAgentic IAM Day 2026 (opens in a new tab)
PlainID

Blog

AI Agent Authorization: Controlling What Agents Can Do at Runtime

Hila Shitrit Nissim · Nov 19, 2026

AI Agent Authorization: Controlling What Agents Can Do at Runtime

At 9:02 a buyer at a manufacturing company launches a procurement agent to reconcile last month's supplier invoices. The agent is registered in Microsoft Entra Agent ID, signs in, and receives an OAuth token with two scopes: read and write on the ERP's vendor module. At 9:11 the agent opens a supplier email that contains a hidden instruction to update the supplier's bank details. The token allows writes to vendor records, so the change goes through.

The agent had a verified identity and valid credentials. It lacked a decision, made at 9:11 instead of 9:02, about whether this specific action fit this agent, this user, and this task. That decision is AI agent authorization, and most enterprises still make it once, at token issuance, then trust the agent with everything the token allows for the rest of its run.

PlainID treats AI agent authorization as a runtime question. Every action an agent takes, from a database query to an MCP tool call, gets evaluated against policy at the moment it happens, using the agent's identity, the identity of the person it serves, the resource, and the purpose of the request.

What AI Agent Authorization Means

AI agent authorization is the process of deciding which data, tools, and actions an AI agent may access on each request, based on the agent's identity, the user it acts for, the resource involved, and the context of the task. It follows authentication, which only proves who the agent is. Authorization governs what the agent does next.

A single prompt can trigger dozens of retrievals and tool calls across systems owned by different teams, and each one is an access decision. Authorize the first step alone and you have authorized all of them.

AI Agent Identity: The Input Every Decision Depends On

The OpenID Foundation's whitepaper on AI agent identity frames the core problem: agents act with some autonomy, often on behalf of a person, and existing identity standards were built for either humans or static services.

In practice you will run three identity patterns. Some agents act for a user, such as a copilot that drafts replies from a sales rep's inbox. Some act alone, such as a nightly reconciliation agent with no human in the loop. Many act in chains, where an orchestrator agent hands work to a research agent that calls a data tool. A single static service account cannot represent any of them well.

Agents should carry their own identity, separate from the people who launch them. An agent registry such as Microsoft Entra Agent ID gives each agent a verifiable identity, so a policy engine can confirm which agent is calling and block one agent from impersonating another. You can see how PlainID builds on that registry in its write-up on extending trust through Microsoft agentic workflows.

Why OAuth Scopes Fall Short for AI Agents

OAuth gives agents a sound way to obtain credentials, the MCP specification builds on OAuth 2.1, and token exchange (RFC 8693) lets a token record that an agent acts for a user. These standards cover how an agent gets access and leave the harder question open.

Scopes are coarse and fixed at the moment of issue. "vendor.write" cannot express "update invoice status but never bank details," and it cannot express "only for suppliers assigned to this buyer." Once issued, a token says nothing about the prompt the agent received, the record it is about to touch, or whether the user's role changed five minutes ago. Developers compensate with wide scopes, and the agent ends up with standing privileges that outlast the work. You can read more on why legacy IAM breaks down once agents enter the stack.

How Runtime AI Agent Authorization Works

Runtime authorization separates the decision from the credential. The agent keeps its token for authentication, and each request passes through an enforcement point that asks a policy engine for a decision before the action runs. The OWASP AI Agent Security Cheat Sheet recommends the same structure: enforce authorization in a component outside the agent's context, and fail closed for unknown tools.

The policy engine sees the agent's attributes (purpose, owning team, risk tier), the user's attributes from your identity provider, the resource's classification, the specific action, and the intent of the request. That last input is what intent-based access control adds on top of policy-based access control, so "reconcile invoices" and "change payment details" produce different answers from the same agent.

Composite Identity Authorization

For an agent acting on a person's behalf, PlainID evaluates both identities as one bound identity and returns a single decision. A CFO using an HR agent should not see every employee's compensation through it. Composite identity authorization computes that intersection on each request and carries it through agent-to-agent and agent-to-MCP hops, so privileges cannot grow as work passes between agents.

Attribute-Based Policies for Ephemeral Agents

Short-lived agents never appear in an access review. PlainID lets you write policies against agent attributes instead, so a temporary agent receives the rights its characteristics justify and nothing beyond them. Agent-only policies cover the cases with no human in the loop.

Continuous Evaluation

PlainID keeps authorizing for the whole session. It evaluates each request with millisecond latency and can stop an agent midway through a task if context changes, which is how runtime authorization delivers zero standing privileges for agentic AI. Policy updates take effect in under 60 seconds.

Where Enforcement Sits in an Agent Workflow

PlainID places enforcement at four points in the agent flow, known as Full-Flow AI Guardrails: the input prompt, data retrieval, MCP tool calls, and the output. In the procurement example, the tool guardrail would have blocked the bank-detail update because no policy permits a reconciliation agent to modify payment fields.

On the tool side, PlainID discovers MCP tools and classifies them by business category, so you write policy against categories such as "payment operations" and new tools fall under existing rules. The PlainID MCP gateway runs inside your environment to enforce those decisions; more detail is in how PlainID secures MCP tool access for AI agents. On the data side, the SQL Authorizer applies row- and column-level policy to agent queries, and pre-retrieval metadata filters restrict what vector and document retrieval can return. Alongside LangChain and LangGraph libraries for agent code, PlainID deploys Authorizers onto existing API and AI gateways (such as Apigee, Kong, AWS, Azure, and Envoy), enforcing runtime policy at the network boundary without modifying downstream microservices.

Choosing an AI Agent Authorization Approach

Most teams start with one of three options. Hardcoded checks inside each agent work for a prototype and turn into shadow policy at twenty agents. Identity provider features handle registration and token issuance, then stop at the token. A dedicated authorization platform manages policy in one place and enforces it at runtime across agents, APIs, data, and applications.

The third option matters once agents share data with the rest of your stack, since one policy for customer PII should govern an agent, an API gateway, and a Snowflake query alike. The PlainID Runtime Authorization Platform governs human, non-human, and AI agent identities under one policy framework, and the Authorization Graph records every decision against the policy that produced it, from identity provider to agent to tool to data. KuppingerCole names PlainID a Product Leader and Market Leader in Policy-Based Access Management, and the platform secures over 35 million identities for Fortune 500 organizations.

Frequently Asked Questions

Should AI agents have their own identity?

Yes. Each agent should carry a distinct, verifiable identity, registered in a system such as Microsoft Entra Agent ID, so policies can tell agents apart and audit trails show which agent acted. For agents that work on a user's behalf, authorization should then evaluate both identities together.

How is AI agent authorization different from authentication?

Authentication proves who the agent and its user are, through tokens from an identity provider such as Okta or Entra ID. Authorization, handled by a runtime platform such as PlainID, decides what the agent may do on each request.

What authorization capabilities do AI agents need?

Agents need per-request decisions instead of static scopes, composite evaluation of agent and user identity, policies based on attributes and intent, enforcement at data and tool level, continuous evaluation through the session, and an audit record for every decision.

Related articles