MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base of adversary tactics and techniques against AI and machine learning systems. Modelled on ATT&CK, it covers behaviours such as data poisoning, model evasion, model stealing, and prompt injection.

How is ATLAS different from ATT&CK?

ATT&CK targets conventional IT systems. ATLAS targets machine learning systems, treating the model, its training data, and its inference interface as attack surfaces. It adds AI-specific tactics such as ML Model Access and ML Attack Staging.

How does ATLAS connect to the EU AI Act and NIST AI RMF?

ATLAS supplies the adversarial threat vocabulary, while the EU AI Act and NIST AI RMF impose obligations to make AI systems robust and secure. Bidda crosswalks each ATLAS technique to those obligations so AI builders can show how a specific threat is governed.

MITRE ATLAS: The Threat Matrix for AI and Machine Learning…

Adversarial ML tactics, from data poisoning to prompt injection, mapped to the EU AI Act and NIST AI RMF.

What ATLAS Covers

ATLAS stands for Adversarial Threat Landscape for Artificial-Intelligence Systems. It extends the ATT&CK model to the machine learning lifecycle, capturing behaviours that traditional security frameworks never anticipated. Where ATT&CK assumes conventional software, ATLAS recognises that the model itself, its training data, and its inference interface are attack surfaces. It is maintained by MITRE and grounded in documented, real-world case studies as well as academic research.

Tactics Specific to Machine Learning

ATLAS reuses familiar tactics such as Reconnaissance, Resource Development, Initial Access, Execution, Persistence, Defense Evasion, Discovery, Collection, Exfiltration, and Impact, and adds tactics unique to AI. ML Model Access describes the levels of access an adversary can obtain to a model, from a query-only API to full white-box access. ML Attack Staging covers the preparation an adversary performs against a model before launching an attack, such as crafting adversarial inputs or building a proxy model.

Representative Techniques

The techniques read like a catalogue of modern AI failure modes: data poisoning to corrupt a model during training, model evasion using adversarial perturbations, model stealing through repeated inference queries, membership inference to determine whether a record was in the training set, prompt injection against large language models, and jailbreaks that bypass safety guardrails. Each is a distinct, documented behaviour with defensive implications.

Why ATLAS Now Has Legal Weight

AI security is no longer purely a research concern. The EU AI Act requires high-risk AI systems to achieve appropriate levels of accuracy, robustness, and cybersecurity, and to be resilient against attempts to manipulate them. The NIST AI Risk Management Framework asks organisations to identify and manage exactly these adversarial risks. ATLAS provides the threat vocabulary; the regulations provide the obligation. The gap between them is what most teams struggle to bridge.

How Bidda Maps ATLAS

Bidda represents ATLAS techniques as verified nodes and crosswalks them to AI governance frameworks: the EU AI Act robustness and cybersecurity provisions, the NIST AI RMF functions, ISO/IEC 42001 for AI management systems, and the OWASP guidance for machine learning and large language model risks. A team building a high-risk AI system can take a technique such as data poisoning and immediately see the governance obligations and control measures that a regulator would expect to be in place.

Built for AI Agents to Query

Because Bidda is agent-native, an autonomous system can consult ATLAS nodes at inference time. The MCP get_mitre_mapping tool resolves an ATLAS technique into its cross-framework view, and the dependency graph lets an agent assemble a complete AI-risk picture. The official ATLAS matrix is published by MITRE at atlas.mitre.org; Bidda adds the verified mapping to enforceable AI governance.

MITRE ATLAS: The Threat Matrix for AI and Machine Learning Systems