MITRE ATLAS

MITRE ATLAS — Adversarial Threat Landscape for AI Systems, a knowledge base of adversarial tactics and techniques targeting AI.

Overview

MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base of adversarial tactics, techniques, and case studies for machine learning systems. Modeled after the widely-used MITRE ATT&CK framework for cybersecurity, ATLAS provides a structured approach to understanding how AI systems can be attacked.

ATLAS helps security teams, red-teamers, and AI engineers identify and defend against real-world adversarial threats to machine learning models and AI applications.

Tactics

ATLAS organizes adversarial behavior into tactics — the “why” behind an attack:

Tactic	Description
Reconnaissance	Gathering information about the target ML system
Resource Development	Acquiring resources for the attack (datasets, models, infrastructure)
Initial Access	Gaining initial access to the ML system or its components
ML Model Access	Obtaining access to the target model (API access, model extraction)
Execution	Running adversarial techniques against the ML system
Persistence	Maintaining access or influence over the ML system
Defense Evasion	Avoiding detection by security controls and monitoring
Discovery	Learning about the ML system’s architecture and behavior
Collection	Gathering data from the ML system (model outputs, training data)
ML Attack Staging	Preparing attack payloads (adversarial examples, poisoned data)
Exfiltration	Extracting data or model information from the target system
Impact	Disrupting, degrading, or destroying the ML system’s function

Key techniques

Technique	Description
Adversarial examples	Crafted inputs that cause the model to misclassify or produce incorrect outputs
Data poisoning	Contaminating training data to introduce backdoors or biases
Model extraction	Querying a model to reconstruct a functionally equivalent copy
Model inversion	Recovering training data or sensitive information from model outputs
Prompt injection	Manipulating LLM behavior through crafted inputs
Backdoor attacks	Embedding hidden triggers in models that activate under specific conditions
Membership inference	Determining whether specific data was used in model training
Model evasion	Crafting inputs specifically to bypass model-based security controls

How Know Your AI maps to MITRE ATLAS

ATLAS Tactic / Technique	Know Your AI Coverage
Reconnaissance / Discovery	System prompt extraction datasets
ML Model Access	API & website evaluation modes
Execution	Red-team attack datasets (15+ methods)
ML Attack Staging	Curated attack datasets in the Marketplace
Adversarial examples	Jailbreak, CIPHER, DAN, and other evasion methods
Data poisoning / Bias	Bias detection datasets
Model extraction	Data extraction attack datasets
Prompt injection	Prompt injection datasets (PAIR, ADAPTIVE, etc.)
Defense evasion	Multi-method attack testing to find guardrail gaps
Exfiltration	PII leakage and data extraction testing
Impact	Security scoring and compliance analysis
Continuous monitoring	SDK monitoring and tracing for production detection

ATLAS vs. ATT&CK

Aspect	MITRE ATT&CK	MITRE ATLAS
Focus	Traditional IT systems	AI and ML systems
Targets	Networks, endpoints, cloud	Models, training pipelines, inference APIs
Techniques	Malware, exploits, phishing	Adversarial examples, prompt injection, data poisoning
Adoption	Industry standard for security operations	Growing adoption for AI security

Resources

• MITRE ATLAS Website
• ATLAS Navigator
• ATLAS Case Studies

Datasets

Browse datasets aligned with ATLAS techniques.

Firewall

Real-time defense against adversarial attacks.