Glossary
MITRE ATLAS (Adversarial ML Threat Matrix)
Definition: MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base of adversary tactics and techniques used specifically against machine learning and AI systems, the AI-specific counterpart to the better-known MITRE ATT&CK matrix.
ATT&CK maps how attackers move through conventional infrastructure. ATLAS maps how attackers manipulate the model itself: prompt injection, jailbreaking, data poisoning, model extraction. The clearest real case: a state-sponsored group (Anthropic’s own disclosure names it GTG-1002) jailbroke Claude Code by claiming, inside the prompt, to be a legitimate security firm running an authorised penetration test. The agent ran an estimated 80-90% of the resulting campaign autonomously across roughly 30 targets. That maps to two specific ATLAS techniques: AML.T0051 (Prompt Injection) and AML.T0054 (LLM Jailbreak). Knowing the technique ID matters because it tells you which control actually closes the gap: a policy layer that verifies claimed authorisation against a source the agent cannot itself modify, not a stronger prompt.