MITRE ATLAS

MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base of the real-world tactics and techniques attackers use against AI and machine-learning systems. Modeled on MITRE's widely used ATT&CK framework, it gives defenders a shared map of how AI systems actually get attacked.

In Depth

MITRE, a nonprofit that runs federally funded research, is best known in security for ATT&CK, a catalog of the tactics and techniques used in cyberattacks. ATLAS applies the same idea to AI. It organizes adversary behavior against machine-learning systems into tactics (the attacker's goals, such as reconnaissance or evasion) and techniques (how they achieve them), grounded in real incidents and published research.

The techniques span the AI lifecycle: poisoning training data, evading a model at inference, extracting or stealing a model, and, increasingly, attacks specific to generative systems like prompt injection and jailbreaks. Because it is structured as a common taxonomy, ATLAS lets security teams describe AI threats in a shared language, map their defenses to specific techniques, and see the gaps.

ATLAS is descriptive rather than prescriptive. It does not tell you how to govern AI or which control to buy. It tells you how AI gets attacked, which makes it a natural companion to a governance framework like the NIST AI RMF and a threat list like the OWASP Top 10 for LLM Applications. Red teams use it to structure their attacks, and defenders use it to check their coverage.

What It Looks Like

A vendor's security team plans its red-teaming against an agent using ATLAS as the map. It works through the relevant techniques (data poisoning, model evasion, extraction, prompt injection) and records which ones the agent resists and which it does not. The result is concrete: coverage mapped to a recognized taxonomy that an enterprise buyer's security team can read and trust, instead of a vague "we tested it."

Why It Matters For AI Vendors

ATLAS turns "we did some security testing" into a defensible, mapped claim about which attack techniques an agent was tested against. Specificity like that holds up in a security review, where a named taxonomy beats a general assurance. It also helps a vendor find blind spots before an attacker or a buyer does, and the techniques it catalogs map to the failures (leakage, manipulation, evasion) that turn into AI liability claims.

Common Questions

ATT&CK catalogs tactics and techniques for conventional cyberattacks. ATLAS applies the same structure to attacks on AI and machine-learning systems, including generative-AI threats like prompt injection and jailbreaks.
No. It is a knowledge base of adversary techniques, not a regulation or certification. Teams use it to structure red-teaming and to check that their defenses cover the known ways AI systems get attacked.
← PreviousLimitations of Liability Next →Model Card

See where your AI agents stand.

Get an Agent Trust Score, map your liability exposure, and find out what it takes to make your AI agents insurable.