⭐ Star on GitHub

Summary: 16 tactics in attack-chain order, from Reconnaissance to Impact, with two AI-specific tactics that have no ATT&CK equivalent.

Sources: raw/MITRE-ATLAS/index.md and the 16 AML.TA* tactic cards (release 2026.06)

Last updated: 2026-07-28


The 16 tactics

In matrix order, with top-level technique counts:

#TacticIDTechniques
1ReconnaissanceAML.TA00028
2Resource DevelopmentAML.TA000313
3Initial AccessAML.TA00047
4AI Model AccessAML.TA00004
5ExecutionAML.TA00056
6PersistenceAML.TA00069
7Privilege EscalationAML.TA00124
8Defense EvasionAML.TA000715
9Credential AccessAML.TA00137
10DiscoveryAML.TA00089
11Lateral MovementAML.TA00152
12CollectionAML.TA00094
13AI Attack StagingAML.TA00016
14Command and ControlAML.TA00144
15ExfiltrationAML.TA00106
16ImpactAML.TA00119

Column counts sum to 113, above the 103 distinct top-level techniques — ten techniques serve more than one tactic and appear in multiple columns, exactly as in ATT&CK.

The two AI-native tactics

Fourteen of the sixteen are ATT&CK tactics carried over. Two are new, and they are where the AI-specific threat model actually lives:

AI Model Access (AML.TA0000) — sits at position 4, between Initial Access and Execution. It encodes the insight that access to the model is a distinct objective from access to the system. An attacker may hold no infrastructure foothold at all and still have what they need, because the API is the attack surface. This is the tactic that makes model extraction legible as an attack chain rather than as API misuse.

AI Attack Staging (AML.TA0001) — position 13, after Collection. Building proxy models, crafting adversarial data, verifying attacks offline before firing them at the target. It has no conventional analogue because conventional exploits do not require you to train something first.

Defense Evasion is the largest column at 15 techniques — consistent with PITAX, where evasion of safety measures is the single biggest theme.

Sub-techniques

70 of the 173 techniques are sub-techniques, using dotted IDs (AML.T0051.001) and linked by specializes relationships. There is no subtechnique-of field — parentage is expressed only through the relationship graph, which is a common source of conversion bugs.

The most consequential example for this vault:

  • AML.T0051 LLM Prompt Injection.000 Direct · .001 Indirect · .002 Triggered

ATLAS splits prompt injection on exactly the axis PITAX uses for delivery, and adds a third for trigger-gated payloads that PITAX handles as a separate technique (T44). See atlas-and-pitax.

Reading the matrix against this vault

Three columns carry most of the vault’s existing material:

Resource Development (13 techniques) is where data poisoning sits — the adversary builds capability before touching the target. That placement is itself the point: poisoning is preparation, and its effects surface much later, which is why detection lags run 6–24 months.