THREATNOIR
Subscribe
Attacker knowledge bases

MITRE ATLAS

A knowledge base of attacks on AI and machine learning systems: prompt injection, jailbreaks, data poisoning and more.

Official source Published by MITRE. Summaries in our own words.

What it is

MITRE ATLAS is a knowledge base of tactics and techniques that adversaries use against AI and machine learning systems, built on real incidents and research. It extends the ATT&CK approach to threats such as prompt injection, model evasion, data poisoning and model theft.

Who it applies to

  • Teams building or deploying machine learning and generative AI systems
  • Security teams threat modelling and red teaming AI applications
  • AI governance and risk functions, for example those preparing for the EU AI Act
  • Researchers and vendors describing AI attacks in a shared language
Status
A free knowledge base maintained by MITRE with industry and government contributors. It is not a standard or law and is updated as new attacks are documented.
Enforcement
Voluntary: there is no certification. It is used as a reference for AI threat modelling and red teaming, and supports risk management under frameworks such as the NIST AI RMF and ISO/IEC 42001.

Key points

  • Many AI attacks reuse normal IT techniques, so ATLAS is meant to be used together with ATT&CK
  • Prompt injection and jailbreaks target generative AI applications and the tools they can call
  • Data poisoning and backdoors target training data and models before deployment
  • Model extraction and inference attacks can leak the model or its training data
  • Case studies are useful for convincing stakeholders that AI threats are real

Where to start

  1. 1List your AI systems, their data sources and what tools or data they can reach
  2. 2Threat model one important system using ATLAS techniques
  3. 3Plan AI red team tests based on the techniques that apply
  4. 4Add relevant mitigations to your AI development and procurement requirements

How it is structured

MatrixTactics and techniques in an ATT&CK style layout, including tactics specific to machine learning
Case studiesDocumented real world and research attacks on AI systems
MitigationsDefensive measures mapped to techniques

A plain-language overview to help you learn, written by ThreatNoir. The official source prevails; this is not legal advice.

Items explained (5)

Ordered by how often our lessons cite them.