Quick answer
MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a public knowledge base of adversary tactics, techniques and real-world case studies for attacks on AI and machine learning systems. It is modelled on MITRE ATT&CK and is used to plan AI red team exercises and track defensive coverage.
On this page
What ATLAS Contains
ATLAS organises attacks on AI systems into tactics (the attacker's goal at each stage) and techniques (how they achieve it), much like ATT&CK does for traditional intrusions. It covers activities such as reconnaissance of AI systems, gaining access to models, prompt injection, data poisoning, model evasion and model or data exfiltration. Each technique links to mitigations and to case studies of real incidents and research.
ATLAS and ATT&CK Together
Most real attacks on AI systems combine both. An attacker may use classic ATT&CK techniques, such as phishing or credential theft, to reach an environment, then ATLAS techniques to manipulate the AI inside it. Testing and detection programs should use both frameworks side by side.
How to Use It
- Plan AI red team exercises around the techniques most relevant to your systems.
- Map findings from AI testing so they can be tracked and compared over time.
- Assess coverage by checking which techniques your monitoring would detect.
- Brief leadership with a recognised framework rather than ad hoc descriptions.
See also the OWASP Top 10 for LLM Applications, which focuses on application vulnerabilities rather than adversary behaviour.
Regulatory and Framework Context
For Australian organisations, ATLAS mapping complements ATT&CK-based reporting already used in red team and SOC programs.
Latest intel on this topic
Updated hourly from the daily briefing.
Common questions
Is MITRE ATLAS free?
Yes. It is publicly available from MITRE, like ATT&CK.
What is the difference between ATLAS and the OWASP LLM Top 10?
ATLAS describes how adversaries attack AI systems across an intrusion; the OWASP list describes the most critical vulnerabilities in LLM applications. They are complementary.
Change log
- 13 November 2026First published.