What it covers. MITRE ATLAS is a public knowledge base of how attackers target AI and machine learning systems. Like ATT&CK, it organises attacker behaviour into tactics and techniques, such as prompt injection, jailbreaking a language model, poisoning training data or stealing a model, and backs them with real case studies.
In practice. Teams that build or deploy AI use ATLAS to threat-model their systems: which techniques apply to a chatbot, an AI agent with tool access or a model trained on outside data, and which controls cover them. Useful evidence is a threat model that names the relevant techniques, tests that try them (for example red-team prompts against guardrails), and monitoring for the signs they leave, such as unusual prompts or tool calls.
Common gaps. Organisations often treat AI features as ordinary software and skip AI-specific threats, or test only the model and not the tools and data it can reach. Prompt injection through documents and web pages the model reads is a frequent blind spot.