The ATLAS Matrix is part of MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems), a knowledge base that documents how adversaries can attack AI and machine-learning (ML) systems across their full lifecycle. It applies the same structured, adversary-centric approach used by MITRE ATT&CK, but is purpose-built for AI systems.
The matrix organizes real-world and theoretical AI attack techniques into a set of tactics that represent an attacker’s objectives, and techniques that describe how those objectives are achieved.
Purpose
The ATLAS Matrix helps organizations:
- Understand how AI systems can be attacked
- Assess risk across AI development and deployment stages
- Design security controls specific to AI and ML threats
- Align AI security with existing threat modeling and SOC workflows
Structure
The matrix is organized into tactics that map to the AI system lifecycle, including:
- Reconnaissance – Gathering information about data, models, or pipelines
- Resource Development – Acquiring data, tools, or infrastructure to attack AI
- Initial Access – Gaining entry into AI systems, data pipelines, or model APIs
- ML Model Access – Interacting with or extracting information from models
- Execution – Running malicious code or adversarial inputs
- Persistence – Maintaining long-term influence over models or data
- Privilege Escalation – Gaining higher levels of access within AI infrastructure
- Defense Evasion – Avoiding detection by AI or traditional security controls
- Credential Access – Stealing secrets used by AI services or pipelines
- Discovery – Mapping model behavior, parameters, or training data
- Collection – Extracting data, models, or sensitive outputs
- Impact – Degrading model integrity, availability, or trustworthiness
Each tactic contains techniques such as data poisoning, model inversion, adversarial examples, prompt injection, and model theft.
Key Characteristics
- Focused on AI-specific threats, not general IT attacks
- Covers training, deployment, and inference phases
- Maps attacks to defensive mitigations and controls
- Designed to integrate with ATT&CK-based threat modeling
Who Should Use It
- Security architects securing AI platforms
- ML engineers and data scientists
- SOC and threat intelligence teams
- Risk and governance teams assessing AI exposure
Why It Matters
As AI systems become embedded in critical business processes, traditional security models are insufficient. The ATLAS Matrix provides a common language and framework to systematically identify, measure, and defend against threats unique to AI.