When teams evaluate an AI agent, attention often goes straight to the model: which provider it uses, how capable it is, and whether its responses are safe.
Those questions matter. But an agent's practical impact also depends on what it can reach and what it is allowed to do.
An agent connected to email, file storage, customer records, source code, and administrative APIs has a different risk profile from an agent that can only answer questions from a public knowledge base. The model may be similar; the available authority is not.
Start with the agent's effective access
To understand an agent's exposure, map the full path from identity to action:
- Identity: Which user, service principal, managed identity, or credential does the agent use?
- Tools: Which functions, connectors, APIs, and integrations can it invoke?
- Permissions: Can it read, write, delete, approve, deploy, or administer?
- Data: Which records, documents, repositories, or databases are reachable?
- Execution context: Where does the agent run, and which network resources can it contact?
- Human oversight: Which actions require confirmation, approval, or a second check?
The configured permissions are important, but effective access may also depend on inherited roles, delegated user access, application logic, and downstream services.
Read access and write access are not the same
A read-only agent can still expose confidential information if it retrieves data for the wrong person or sends it to an inappropriate destination. Write access adds other possibilities: changing records, sending messages, modifying code, or triggering business workflows.
The risk is therefore not captured by a simple count of connected tools. Teams need to understand what each integration permits and what consequences follow from an action.
For example, a support agent may need to read a customer's order status. It may not need permission to export every customer record, alter account ownership, or issue refunds without approval. Narrowing the available operations reduces the impact of mistakes and misuse.
Apply least privilege to agents
The principle of least privilege is straightforward: give an identity only the access needed for its defined task, for only as long as needed where practical.
For AI agents, apply it at several layers.
Use dedicated identities. Avoid sharing broad human credentials with agents. Dedicated identities make ownership, access review, and revocation easier to manage.
Limit tool capabilities. Prefer a narrowly scoped function such as “retrieve order status” over a generic tool that can execute arbitrary queries or actions.
Separate reading from acting. Where possible, use distinct permissions or workflows for retrieving information and making consequential changes.
Require approval for high-impact actions. Sending external communications, changing access, moving money, deleting data, or deploying code may warrant confirmation or a controlled workflow.
Review access regularly. Remove unused integrations, rotate credentials according to policy, and reassess permissions when an agent's purpose changes.
Include sensitive data in the review
An agent's data exposure can be broader than its stated purpose. Review whether connected sources contain personal information, health information, financial data, credentials, source code, or confidential business records.
Do not assume that an agent processes sensitive data simply because it has a connection to a system that contains it. Verify the available permissions and data paths. Likewise, do not assume data is safe because the agent's user interface does not display it; logs, tool outputs, retrieval layers, and downstream services may handle information differently.
Build an access map, not just a tool list
For each agent, record its identity, connected tools, permission scope, data sources, action capabilities, owner, and evidence source. Identify high-impact paths—for example, an agent that can retrieve confidential documents and send messages externally.
This map helps security teams ask better questions: Is every connection necessary? Can a compromised or manipulated agent reach a critical system? Are actions logged? Can access be revoked quickly?
Treat access as a lifecycle
Permissions change as agents evolve. A new connector, tool, model, or workflow can materially change an agent's risk. Include access review in development, deployment, and change management, and investigate unexpected permission changes.
An agent's intelligence does not determine its authority. Its identity, tools, and permissions do.
To understand the risk an agent introduces, look beyond what it says—and examine what it can actually access and do.
