Quick answer: MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base that gathers real world attack tactics and techniques against artificial intelligence and machine learning systems. It is the AI adapted version of the classic MITRE ATT&CK. It classifies AI specific threats, from data poisoning to model theft, from evasion attacks to prompt injection, as tactics and techniques. For defenders it provides a common language for threat modeling AI systems.

What is MITRE ATLAS used for?

Traditional security frameworks do not cover AI specific attacks. Poisoning a model's training data, fooling it at inference time, or stealing it through an API does not map neatly onto the classic ATT&CK matrix. ATLAS fills exactly this gap.

ATLAS organizes attacks against AI systems with the same logic as ATT&CK: tactics (the attacker's goal) and techniques (the way to reach that goal). This lets an organization speak a common, standard language about which attacks are possible when threat modeling an AI system.

The difference between ATLAS and ATT&CK

Topic MITRE ATT&CK MITRE ATLAS
Scope Traditional IT systems Artificial intelligence and ML systems
Example tactic Initial access, lateral movement ML model access, ML attack staging
Example technique Phishing, credential theft Data poisoning, evasion, model extraction
Purpose IT threat modeling AI threat modeling

ATLAS is not standalone, it is aligned with and complements ATT&CK. An attack on an AI system often uses both classic IT techniques (accessing the server) and AI specific techniques (fooling the model) together.

Main technique groups in ATLAS

  • Reconnaissance: Learning the model's type, API and weak points.
  • Data poisoning: Corrupting the model by injecting malicious samples into training data.
  • Evasion: Misleading the model by altering the input at inference time.
  • Model extraction and theft: Recovering the model or training data through API queries.
  • Prompt injection: Hijacking instructions in LLM based systems.

The KAOS and DSET approach

DSET uses MITRE ATLAS as a threat modeling framework in the security review of AI systems. When evaluating an AI application, we map ATLAS tactics and techniques to the target's real attack surface. Our KAOS security engine tests ATLAS techniques such as prompt injection and data leakage in AI applications in an evidence based way. Our goal is not a theoretical list but showing which attack is actually possible.

Related reading: MITRE ATT&CK framework, prompt injection and OWASP LLM Top 10, AI data poisoning, AI model security and model theft, AI red teaming methodology, adversarial ML and evasion attacks.

Frequently asked questions

What is the difference between MITRE ATLAS and ATT&CK? ATT&CK covers attack tactics and techniques against traditional IT systems. ATLAS adapts the same logic to artificial intelligence and machine learning systems: it classifies AI specific techniques such as data poisoning, evasion, model theft and prompt injection. They are aligned, and an AI attack often uses both together.

Who uses ATLAS? ATLAS is used by the security teams of organizations that build and operate AI systems, by AI red teams and by auditors. It provides a common language and reference for threat modeling AI systems, so that which attacks are possible is discussed in a standard way.

Is ATLAS a tool or a framework? ATLAS is not a software tool but a knowledge base and framework. It documents real world attacks against AI systems as tactics and techniques. It is used as a reference in threat modeling, red team planning and defense prioritization.

Sources

For MITRE ATLAS based threat modeling and security review of your AI systems, you can talk to our team at Ankara Hacettepe Technopark. Contact DSET.