Skip to content
be-human.ai
← Back to the journal

Human oversight

AI Agent Attack Surface: Map Every Entry Point Before Review

A practical method to list every channel, tool and credential an AI agent touches, so security review keeps pace as the agent's reach grows.

Be Human typographic cover with the title: AI Agent Attack Surface: Map Every Entry Point Before Review
Conceptual illustration created with AI · Be Human

The AI agent attack surface is the full set of inputs, actions and credentials through which an agent can be influenced or made to act. To map it, list every channel the agent reads from, every tool it can call and every identity those calls run under, then assign an owner and a review trigger to each line. The result turns a general worry about a growing agent into a finite list that a security review can check.

Key takeaways

  • The attack surface of an agent is defined by what it can read, what it can do and which identities let it do those things.
  • Build three inventories, channels, tools and credentials, then join them task by task.
  • Every new tool, connector or shared mailbox is a change to the map and should pass review before it goes live.
  • Any text that reaches the agent through a channel may contain instructions, so treat inbound content as untrusted.
  • Review the map on a fixed schedule and after every incident, not only at launch.

Why a growing agent needs a map instead of a launch review

A security review done before launch describes the agent as it was on the day it went live. Agents rarely stay that size. A support assistant often starts with a knowledge base and a ticket queue, then gains access to a calendar, a billing view and a shared inbox, because each addition looked reasonable when it was requested. Each addition widens the surface a little, and no single change looks dangerous enough to stop.

A map makes that reach visible. It is a living list of every place the agent reads from, every action it can take and every secret it uses to do either. Once the list exists, review becomes a comparison: does the current agent match what the team approved? Differences are the points that need a decision, and they are much easier to find than a vague sense that the agent has grown.

Step 1: Inventory every channel the agent reads from

Start with inputs, because the attack surface begins wherever outside text or data can reach the agent. Prompt injection is the best known case: instructions hidden in a web page, an email signature or an uploaded document can try to redirect the agent's behavior. The OWASP prompt injection prevention guidance is a useful reference for this class of problem. The working principle is simple: content from any channel may contain instructions, whether or not the sender meant to include them.

For each channel, record the following:

  • Who can write into it, such as customers, staff, vendors, the public or other agents.
  • Whether content comes from a known person or arrives automatically through a form, a webhook or a scraped page.
  • Whether the agent sees the full content or only a summary produced by another system.
  • Whether the channel is shared with other systems that could alter it.
  • Whether anything the agent reads is stored in its memory afterward.

Include indirect channels as well. Retrieved documents, notes the agent wrote into memory last week and outputs from other agents are all inputs. Teams often list the obvious inbox and forget the knowledge base the agent searches quietly every time it answers a question. For a closer look at inbound text risks, read prompt injection risks for AI agents.

Step 2: List every tool and the action it allows

A tool is any function the agent can call: send a message, create a record, update a status, issue a refund, query a database or run a script. Tools are where a read-only assistant becomes an actor, so each one needs its own line in the map. The Model Context Protocol introduction describes how tools and data sources are exposed to models, which gives you useful vocabulary when you ask an engineering team what is connected.

Classify every action by its effect:

  • Read: viewing a record or document with no change to any system.
  • Draft: producing text or a proposed change that a person must release.
  • Write: changing a record, status or file inside a company system.
  • External send: contacting someone outside the company by email, text, chat or voice.
  • Commitment or money: anything that spends funds, changes a price, accepts terms or moves money.
  • Delete or export: removing data or moving it outside its approved store.

The last three categories deserve the strongest approval gates. A tool that looks harmless in a list can still reach them through a parameter. A tool named update note may accept a field that changes a customer's status, so record the parameters that matter, not only the tool name. The entitlement mapping guide for agents shows how to pair each tool with the records and actions it can reach.

Illustrative scenario: a scheduling agent for a services firm has a tool called check availability. The first map marks it as read. A later review finds that the same tool accepts a client identifier and returns contact details, which moves it into the data access category and triggers a privacy check before it stays enabled.

Step 3: Trace every credential and where it lives

Every tool call runs under an identity. That identity may be a service account, an OAuth grant tied to a staff member, an API key in a configuration file or a session token the agent reuses. Credentials belong in the attack surface because they decide what a compromised or misled agent can actually do. The OAuth 2.0 documentation explains how scopes limit what a token can access, and scope is the question to ask about every grant.

For each credential, record:

  • Which tools or channels use it, and under whose identity.
  • Its scope, written in plain words such as read the shared calendar or send mail from the support address.
  • Who issued it and who can revoke it.
  • Its expiry and the last time it was rotated.
  • Where it is stored, and whether it appears in logs, prompts, memory or screenshots.

The last item is the one teams skip. A key that is safe in a secrets store can still appear in a log line or in text the agent copies into memory. If the map shows a credential inside memory, the memory becomes part of the credential review.

Check each grant against least privilege. An agent that drafts replies rarely needs write access to the customer database. The OWASP authorization guidance is a practical reference for confirming that each action is enforced by the system that holds the data, not only suggested in the agent's instructions. When a staff member leaves, their delegated grants must be removed too, a step covered in offboarding staff from agent tools.

Step 4: Join the three lists into a task map

Channels, tools and credentials become useful only when they are joined. A task map follows one job from start to finish, so that a reviewer can see the whole path at once. Work through one recurring task at a time:

  • Pick one recurring task with a clear owner, such as answering billing questions or drafting quotes.
  • Write down the channel that starts it and every input that can shape the output.
  • List each tool call in order, with its action category beside it.
  • Record the credential behind each call and its scope.
  • Mark every step that changes a record, contacts someone outside the company or moves money, and name the person who approves it.
  • Check whether any high-risk step can run without a person reviewing it first.

Illustrative scenario: a small professional services firm lets an agent draft quotes from inbound web forms. Joining the lists shows that the web form is a public channel, the price sheet is a read tool, the draft is a draft action and the send to the client is an external send under a shared sales mailbox credential. The gap is visible immediately: the external send has no approver, and the shared mailbox credential is used by two other tools. The fix is to give the agent its own scoped credential and require a person to release each send.

Keeping the map current after launch

The map is only useful if it changes when the agent changes. Treat the following events as triggers for an update and a short review:

  • A new tool, connector or plugin is added, including one that was only meant for testing.
  • A new channel opens, such as a web form, a phone line or a shared mailbox.
  • A credential is renewed, reissued or shared with another system.
  • A staff member changes role or leaves, and their delegated access is reviewed.
  • An incident, near miss or unexpected output appears in the logs.
  • The agent's memory expands to store a new type of record.

Assign one owner for the map. That owner is usually an operations leader supported by technical staff, not the agent itself. Attach a short note to each change explaining why it was approved. Pair the map with the pre-launch security review checklist so that new capabilities answer the same questions before they go live.

How Be Human handles this: Be Human designs and operates personalized AI agents whose identity, memory, channels, tools, permissions and operating context stay under human control. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, which gives the map a documented starting version. Managed operations then cover monitoring, permission review, integration maintenance and incident response, so the map is updated when reach changes rather than rediscovered after a problem. Once you know which task you want to start with, the free Be Human AI diagnostic will help you find the first task worth handing to an agent.

Frequently asked questions

What is an AI agent attack surface? The AI agent attack surface is the set of inputs, actions and credentials through which an agent can be influenced or made to act. It includes every channel that feeds the agent text or data, every tool it can call and every identity those calls run under. Mapping all three gives a review team a finite list to check.

How often should the map be reviewed? Review it on a fixed schedule that matches how quickly the agent changes, and also whenever one of the triggers in this article occurs. A monthly check of tool and credential changes is a reasonable starting point for many small teams, and the result should be recorded. Any incident or near miss calls for an immediate review of the lines it touches.

Do we need technical staff to build the map? Not to draft it. Operations leaders and business owners can list channels, actions and approvers with the people who built the agent, then ask engineers to confirm credentials, scopes and where secrets are stored. The business side should own the decisions about what the agent may touch, and engineering should confirm how those decisions are enforced.

Is a map enough to make an agent safe? No. A map shows where the risk is, but it does not stop a misled agent on its own. It needs permissions enforced by the systems that hold the data, approval gates for high-risk actions and logs that let people reconstruct what happened.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →