What is an authorization boundary?

An authorization boundary is the point where a system actually checks whether an action is allowed and refuses it if not, such as an IAM role, a database permission or a sandbox.

Share

What is an authorization boundary?

An authorization boundary is the point where a system actually checks whether an action is allowed and refuses it if not, such as an IAM role, a database permission or a sandbox.

Why it matters

AI agents are often limited by instructions in their prompt or by an "approve" button in the app. Neither is a check the target system enforces, so a confused or manipulated agent can still do anything its credentials allow. Knowing where the real boundary sits tells you what an agent can actually do.

Example

On 13 November 2026 Microsoft added write tools to its Azure Resource Manager MCP server, which runs every call with the signed-in user's Azure rights. Microsoft's own governance page says a prompt such as "don't change resources" "isn't an authorization boundary" and that approving a tool call in the client grants no permission (Microsoft Learn).

How to defend against it

  • List every tool an agent can call and write down which system enforces its limit (RBAC, database role, network policy, sandbox).
  • Give agents their own identity with the narrowest role that does the job, instead of the user's full rights.
  • Test the limit: have the agent attempt an out-of-scope action and confirm the target system refuses it.
  • Log actions under the agent's identity so you can see what it did (see the AgentSec checklist).

Delegated authority, excessive agency, confused deputy, least privilege

Sources

From AgentSec Brief, the daily brief on securing agentic AI. Subscribe free at https://www.agentsecbrief.com