OWASP LLM Top 10Agentic Top 10MITRE ATLAS

Threat modeling for AI & ML systems

STRIDE and a classic DFD get you most of the way for the infrastructure around a model -- the API, the storage, the auth. They don't cover prompt injection, training data poisoning, or an agent deciding to call a tool nobody explicitly coded. ThreatTree adds OWASP's LLM and Agentic frameworks and MITRE ATLAS on top, so both layers live in one forest.

  • Free forever plan
  • Covers chatbots and tool-calling agents
  • Export anytime, no lock-in
ThreatTree data flow diagram of a customer support assistant with the AI and ML module on: a support agent, a hosted LLM and a knowledge retriever sit inside a dashed Context Window boundary marked high privilege, with a refund tool, an orders database and a vector store outside it. The hosted LLM is selected with data classification PII in Prompts.
Real screenAI & ML module: a Context Window boundary and an LLM handling PII in prompts. Industry Modules are included on Pro and Enterprise.

What is AI threat modeling?

AI threat modeling is threat modeling of systems that use AI: LLM applications, retrieval-augmented generation, agents and the tools they call. It adds threats ordinary applications don't have, such as prompt injection, data poisoning and an agent taking actions it shouldn't, usually mapped to the OWASP Top 10 for LLM Applications and MITRE ATLAS. It is about modeling AI systems, not about using AI to write the threat model.

OWASP LLM/Agentic Top 10 & MITRE ATLAS

AI-specific threats, mapped to ThreatTree

These frameworks were built specifically for the parts of an AI system a classic DFD-plus-STRIDE pass walks straight past.

What's different about AI systems In ThreatTree
1 Model, orchestrator & retrieval architecture
The LLM, vector store/RAG pipeline, and any training pipeline feeding it
Model each as a DFD process or data store, with a trust boundary around the prompt/context window -- the point where user input directly shapes model behavior
2 Classic infrastructure threats
The API, storage, and auth layers around the model
STRIDE, applied to each component as to any other DFD element -- this layer hasn't stopped needing it
3 LLM-intrinsic threats
Prompt injection, sensitive information disclosure, supply chain, insecure output handling, unbounded consumption
Select OWASP Top 10 for LLM Applications as the Threat Framework on any attack tree node
4 Agentic threats
Excessive agency, unsafe tool invocation, goal manipulation for systems that act autonomously
Select OWASP Top 10 for Agentic Applications alongside or instead of the LLM Top 10, on the same node
5 Adversary tactics & techniques
How an attacker actually executes an AI-specific attack
A MITRE ATLAS technique field appears automatically alongside the existing MITRE ATT&CK field -- no separate tooling required

What this is, and isn't: ThreatTree covers the threat modeling and documentation phase for AI systems -- identifying, decomposing, and scoring threats before or alongside development. It doesn't run dynamic prompt-injection tests or red-team a live model; pair it with your adversarial-testing process for that.

AI-assisted attackers

Attackers use AI too, against systems that have none

Securing your AI system is half the picture. The other half is that attackers now use AI against everything else you run: your helpdesk, your login page, your patch cycle. Mostly it makes attacks you already model cheaper, faster and more convincing. ThreatTree's threat dictionary covers both halves, on every plan.

What AI changes for the attacker Suggested threat in ThreatTree
1 Impersonation
A cloned voice or a deepfake video call passes for an employee or an executive
On Identity Provider: a cloned voice passes helpdesk checks for an MFA or password reset. On External Entity: an executive deepfake asks for a payment. On the SaaS Admin / Support Console: a caller talks support into changing account access.
2 Phishing at scale
Fluent, personal messages in any language, for almost no cost per target
On External Entity: AI-written spear-phishing. On Email / SMS Service: AI-written emails impersonate your brand to your customers.
3 Faster exploitation
Less time between an advisory and a working exploit, and code read at machine speed
On Application Server: AI shortens the time from disclosure to working exploit. On API Gateway: agents test authorization at scale. On Source Code Repository and Mobile App: leaked code and app binaries are mined for flaws and secrets.
4 Automation that looks human
CAPTCHAs solved, bot traffic that varies like people do
On Web Client (SPA): AI solves the CAPTCHA. On API Gateway: bots imitate human traffic to scrape or abuse the API.
5 Your developers' AI as a way in
Hallucinated package names registered by attackers, jailbreaks found by automated search
On CI/CD Pipeline: a coding assistant installs a hallucinated package. On LLM / Model Endpoint: attackers automate jailbreak search with another model.
6 Detection and response
Self-rewriting malware and intrusions run by AI agents, at machine speed
On Logging / Observability: signature-based detection misses self-rewriting malware, and manual alert review is too slow.

How they're marked: each of these threats carries MITRE ATT&CK T1588.007 (Obtain Capabilities: Artificial Intelligence) next to the technique it speeds up, plus ISO 27001, NIST CSF, SOC 2 or NIST AI RMF controls on Pro. They're suggestions you review like any other: Generate attack trees drafts them from your diagram and they stay out of the risk register until you accept or score them. ThreatTree has no built-in AI and doesn't detect AI attacks; it helps you decide which of your risks they change. See the worked example.

Why teams use ThreatTree

One forest, two threat layers

Infrastructure threats and AI-specific threats live on the same architecture, not two disconnected documents.

01

Multi-Framework Threat Tagging

Tag every node with STRIDE, OWASP LLM Top 10, OWASP Agentic Top 10, CAPEC, or MITRE ATT&CK/ATLAS -- see which frameworks your model covers and where the gaps are.

02

Attack Path Decomposition

Break each threat into an Attack Tree with AND/OR logic, from attacker goal to atomic step -- e.g. how an indirect prompt injection in a retrieved document leads to a tool call nobody intended.

03

The Prompt as a Trust Boundary

Draw the context window as its own trust boundary -- anything that ends up in it, from user input to retrieved documents to tool outputs, is treated as the high-privilege surface it actually is.

04

Prioritized Risk Register

Likelihood x Impact scoring generates a ranked risk register automatically across every tree in a forest -- useful evidence for a NIST AI RMF Map/Measure/Manage writeup.

05

The Whole AI Stack in the Palette

Beyond LLM endpoints, agents, tools, retrievers and vector stores, the AI & ML group covers MCP servers, guardrails, model registries, embedding models and evaluation pipelines.

06

Threats Tagged to OWASP and NIST AI RMF

The newer AI elements' suggested threats arrive tagged with their OWASP Top 10 for LLM Applications or Agentic category and mapped to NIST AI RMF, so a generated tree starts with its framework references in place.

FAQ

AI & ML threat modeling, answered

What teams building LLM and agentic systems ask before adopting ThreatTree.

Still have questions?

We usually reply within a day.

Get in touch
Does ThreatTree replace red-teaming or prompt-injection testing?

No. ThreatTree covers the threat modeling and documentation phase -- identifying and scoring threats before you build or ship. It doesn't run dynamic tests against a live model; pair it with your red-team or adversarial-testing process for that.

Does ThreatTree support agentic, tool-calling systems, not just chatbots?

Yes. Alongside OWASP Top 10 for LLM Applications, ThreatTree supports OWASP Top 10 for Agentic Applications as a Threat Framework option on any attack tree node, covering risks like excessive agency and unsafe tool invocation.

Do we need deep ML expertise to use ThreatTree for an AI system?

No. The OWASP LLM/Agentic Top 10 and MITRE ATLAS fields give your team a named framework to work through, the same way STRIDE guides a classic DFD review -- you don't need to invent an AI-specific threat taxonomy from scratch.

Does ThreatTree cover attackers who use AI, not just attacks on AI systems?

Yes. The threat dictionary includes threats from AI-assisted attackers on ordinary elements, such as cloned-voice helpdesk resets on an identity provider, AI-written phishing, AI-solved CAPTCHAs, faster exploitation of public-facing servers and hallucinated packages in a CI/CD pipeline. They're available on every plan, without the AI & ML module, and each one is tagged with MITRE ATT&CK T1588.007.

Where can I see a worked example?

The blog post linked below walks through a realistic internal support agent -- RAG retrieval, an LLM call, and an autonomous ticket-creation tool call -- threat modeled with STRIDE, OWASP LLM/Agentic Top 10, and MITRE ATLAS together.

Working through AI system threat models? Get new threat-modeling guides by email — no more than one a week.

Start your AI system threat model in ThreatTree

Free plan available -- no credit card required. Be up and running in minutes.