START HERE

What you will learn

  • Describe assets, actors and trust boundaries in plain English.
  • Turn a plausible failure into a control and a test.
  • Keep a threat model useful as the system changes.

Before you start: Choose one proposed application or integration. Paper and a simple diagram are enough to start.

Reference edition: OWASP Threat Modeling Cheat Sheet (living guidance). Sources checked 2026-11-11.

Start with the system rather than an attack catalogue

An asset is something you need to protect: a private document, a service’s availability or permission to deploy a release. An actor is a person or system interacting with it. A trust boundary is a point where information or an operation crosses into a different permission context.

For a document assistant, useful boundaries include the user’s request reaching the application, retrieval reaching storage, and a proposed tool action reaching an external service. Drawing those crossings makes it easier to notice assumptions that a single “AI assistant” box would hide.

A threat needs a consequence and a response

“The system may be hacked” is too broad to guide a design change. “A read-only user may receive another customer’s document through retrieval” identifies an actor, an asset, a crossing and an unwanted outcome. You can now discuss what should prevent it and how to test that prevention.

Threat modelling is not a one-off diagram or a claim to have predicted every attack. Revisit it when a feature gains a new data source, role or tool permission. Also check ordinary failures: an unavailable approval service can create risk if the application treats its absence as permission to continue.

Step-by-step process

  1. Describe the intended use

    State the task, actors, protected information and consequential operations. Keep one workflow in scope for the first pass.

  2. Draw the important data and action paths

    Show where information enters, where it is stored and where a command or message can leave the system. Label the permission changes.

  3. Write plausible failure statements

    Use the pattern: “This actor could cause this unwanted outcome through this path.” Include mistakes and unavailable dependencies as well as deliberate misuse.

  4. Choose a response for each important threat

    You might restrict a tool, add an access check, redesign the feature or decide not to support an operation. Record who may accept any risk that remains.

  5. Check the response and maintain the record

    Define an authorised test or review that would reveal whether the control works. Update the threat model when the design or evidence changes.

FOLLOW ONE EXAMPLE

Worked scenario

Adding document search to a stakeholder assistant

A fictional prototype starts with public documents. Its owner proposes adding customer files and an email action, creating new permission boundaries.

  1. The diagram now shows retrieval from private storage and an outbound email operation. These are different capabilities from summarising public text.
  2. The owner records two threats: retrieval returning another customer’s material, and an untrusted document requesting an unauthorised recipient.
  3. The proposed responses are identity-aware retrieval and recipient/approval checks outside the model. Synthetic records and a harmless instruction fixture are used to check those boundaries.

What this shows: The design discussion produces specific controls and tests before live data or new permissions are introduced. Unresolved risks remain visible to the decision-maker.

The process at a glance

The model should change a decision or a test. Keep the connection between the failure, chosen response and evidence visible.
ProcessDecisionHuman checkpoint
Read Mermaid diagram source
flowchart LR
 A[Describe the feature] --> B[Map assets and boundaries]
 B --> C[Identify plausible failures]
 C --> D[Choose controls and owners]
 D --> E[Test the controls]
 E --> F[Update the model when things change]

What you should have at the end

  • A small diagram showing the paths that matter to the feature.
  • A list connecting each important threat to a control, owner and check.
  • A clear reason to revisit the model after a capability change.

Common mistakes and how to avoid them

Listing frightening terms without a system path

Identify how the unwanted outcome could occur in this design.

Drawing a control without checking it

A box labelled “authorisation” is an intention until the behaviour is inspected.

Ignoring normal service failures

Ask what happens when identity, approval or storage services are unavailable.

APPLY THE IDEA

Try it yourself

Your task: Map one integration you own. Identify one protected asset, one boundary and one operation that could have an unwanted effect.

What to produce: Three threat statements with proposed responses, owners and an authorised way to check them.

Check your understanding: Can you explain what would change in the design or test plan because of each threat?

Key terms in plain English

Asset
Information, capability or service whose loss or misuse matters.
Actor
A person or system that interacts with the feature.
Threat
A potential cause of an unwanted outcome.
Trust boundary
A crossing between different permission or trust contexts.

Sources and further reading