MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is the world’s primary threat‑intelligence framework describing how attackers target AI models, data pipelines, training processes, and inference systems. Think of it as the MITRE ATT&CK for AI — but focused specifically on adversarial machine‑learning attacks.
It is maintained by MITRE, the same organisation behind ATT&CK, CVE, CWE, and other global security standards.
A structured knowledge base of:
AI attack techniques
Real‑world adversary behaviours
Documented case studies
Defensive recommendations
Mappings to traditional ATT&CK tactics
It helps organisations understand how AI systems are attacked, how to detect those attacks, and how to defend against them.
AI systems introduce new attack surfaces that traditional cybersecurity frameworks don’t cover:
Model manipulation
Training data poisoning
Prompt injection
Adversarial examples
Model extraction
RAG exploitation
Agentic misuse
MITRE ATLAS provides a shared language and structured model for these threats.
Data poisoning
Label flipping
Backdoored datasets
Supply‑chain compromise of training sources
Model extraction
Model inversion
Adversarial examples
Gradient manipulation
Prompt injection
Jailbreaking
RAG manipulation
Agentic exploitation
Compromised datasets
Malicious pretrained models
Dependency attacks
API abuse
Resource exhaustion
Denial‑of‑service
Unauthorized model access
Organisations use ATLAS for:
Threat modelling AI systems
Detection engineering for AI pipelines
Red teaming LLMs and ML models
Risk assessments aligned to NIST AI RMF
Zero Trust for AI architecture
AI security audits and compliance
It is widely adopted by:
Cloud providers
Financial institutions
Government agencies
AI product companies
Cybersecurity consultancies