DETECT hunt Run and measure

IdeaMeasure

ATLAS and D3FEND

ATT&CK covers attacks on hosts, networks, identities, and cloud services. MITRE ATLAS lists attacks on AI systems, and MITRE D3FEND names the defenses that counter ATT&CK techniques. This page says how the framework uses each, and where its copy of their data differs from the source.

ReferenceFramework v0.136.0. ATLAS 2026.09, D3FEND 1.6.0, ATT&CK 20.1.

On this page

ATLAS

ATLAS 2026.09 lists 16 tactics, 120 techniques, 88 sub-techniques, 40 mitigations, and 73 case studies. Of the case studies, 23 are incidents against deployed AI systems and 50 are exercises by red teams and researchers. ATLAS adopted 44 of its 208 techniques and sub-techniques from ATT&CK and names the ATT&CK technique each came from, as with Exploit Public-Facing Application. The rest are specific to AI, such as prompt injection, jailbreaks, and poisoned training data.

ATLAS lists no log sources and no analytics, so the coverage map can't score its techniques the way it scores ATT&CK's. A score.yml accepts an ATLAS id such as AML.T0051 beside its ATT&CK ids. The kit's lint checks it against the pinned ATLAS release, so a detection for an AI system enters the backlog like any other.

Collect the model's telemetry first

Of the 40 ATLAS mitigations, 5 detect an attack in progress: Predictive AI Adversarial Input Detection (AML.M0015), AI Telemetry Logging (AML.M0024), Deepfake Detection (AML.M0034), AI Agent Scope Drift Detection (AML.M0038), AI Honeypots (AML.M0039). The other 35 prevent, harden, or limit. ATLAS files its mitigations as policy or technical controls, so this split by what each one does is this site's reading. AI Telemetry Logging records the inputs and outputs of a deployed model, and 18 techniques name it. A detection for prompt injection, or for exfiltration through an inference API or an agent's tools, has nothing to read without that log. Treat it like any other log source: give it a row in inventory.csv and a silence alert under DET-0000.

ATT&CK's own AI techniques

ATT&CK 20.1 has 7 techniques of its own for AI: Direct Prompt Injection (T1699), Indirect Prompt Injection (T1698), AI-Enabled Exfiltration (T1700), Impair Autonomous Defenses (T1685.007), AI Result Poisoning (T1608.007), Query Public AI Services (T1682), and Artificial Intelligence (T1588.007). The coverage map scores them like any other technique. ATT&CK names log sources for T1699, T1698, and T1700 only, so the map can't yet say which logs the others need. T1608.007, T1682, and T1588.007 are PRE techniques, so the map starts them out of scope. A detection for a technique both knowledge bases list carries both ids: the ATT&CK id counts toward coverage, and the ATLAS id ties it to ATLAS's case studies and mitigations.

Techniques ATLAS ties to a mitigation that detects

48 of the 208 techniques and sub-techniques name at least one detecting mitigation. The techniques with the most ATLAS case studies come first.

TechniqueTacticsMitigation that detectsCase studiesFrom ATT&CK
AML.T0053 AI Agent Tool InvocationExecution, Privilege Escalation, Lateral MovementAI Telemetry Logging (AML.M0024)19
AML.T0015 Evade AI ModelInitial Access, Defense Evasion, ImpactPredictive AI Adversarial Input Detection (AML.M0015), Deepfake Detection (AML.M0034)18
AML.T0051.001 LLM Prompt Injection: IndirectExecutionAI Telemetry Logging (AML.M0024)18
AML.T0047 AI-Enabled Product or ServiceAI Model AccessAI Telemetry Logging (AML.M0024)16
AML.T0093 Prompt Infiltration via Public-Facing ApplicationInitial Access, PersistenceAI Honeypots (AML.M0039)13
AML.T0051.000 LLM Prompt Injection: DirectExecutionAI Telemetry Logging (AML.M0024)11
AML.T0040 AI Model Inference API AccessAI Model AccessAI Telemetry Logging (AML.M0024), AI Honeypots (AML.M0039)10
AML.T0086 Exfiltration via AI Agent Tool InvocationExfiltrationAI Telemetry Logging (AML.M0024)10
AML.T0000 Search Open Technical DatabasesReconnaissanceAI Honeypots (AML.M0039)8T1596
AML.T0049 Exploit Public-Facing ApplicationInitial AccessAI Honeypots (AML.M0039)8T1190
AML.T0031 Erode AI Model IntegrityImpactPredictive AI Adversarial Input Detection (AML.M0015)6
AML.T0006 Active ScanningReconnaissanceAI Honeypots (AML.M0039)5T1595
AML.T0051.002 LLM Prompt Injection: TriggeredExecutionAI Telemetry Logging (AML.M0024)5
AML.T0085.001 Data from AI Services: AI Agent ToolsCollectionAI Telemetry Logging (AML.M0024)5
AML.T0116 Autonomous ReconnaissanceReconnaissanceAI Agent Scope Drift Detection (AML.M0038)4
AML.T0117 Autonomous Attack-Path AdaptationAI Attack AdaptationAI Agent Scope Drift Detection (AML.M0038)4
AML.T0043.000 Craft Adversarial Data: White-Box OptimizationAI Attack AdaptationPredictive AI Adversarial Input Detection (AML.M0015)3
AML.T0043.002 Craft Adversarial Data: Black-Box TransferAI Attack AdaptationPredictive AI Adversarial Input Detection (AML.M0015)3
AML.T0043.003 Craft Adversarial Data: Manual ModificationAI Attack AdaptationPredictive AI Adversarial Input Detection (AML.M0015)3
AML.T0088 Generate DeepfakesAI Attack AdaptationDeepfake Detection (AML.M0034)3
AML.T0095 Search Open Websites/DomainsReconnaissanceAI Honeypots (AML.M0039)3T1593
AML.T0095.000 Search Open Websites/Domains: Code RepositoriesReconnaissanceAI Honeypots (AML.M0039)3T1593.003
AML.T0101 Data Destruction via AI Agent Tool InvocationImpactAI Telemetry Logging (AML.M0024)3
AML.T0000.003 Search Open Technical Databases: Scan DatabasesReconnaissanceAI Honeypots (AML.M0039)2T1596.005
AML.T0005.001 Create Proxy AI Model: Train Proxy via ReplicationAI Attack AdaptationAI Telemetry Logging (AML.M0024)2
AML.T0029 Denial of AI ServiceImpactPredictive AI Adversarial Input Detection (AML.M0015)2
AML.T0043.001 Craft Adversarial Data: Black-Box OptimizationAI Attack AdaptationPredictive AI Adversarial Input Detection (AML.M0015)2
AML.T0052 PhishingInitial Access, Lateral MovementDeepfake Detection (AML.M0034)2T1566
AML.T0085.000 Data from AI Services: RAG DatabasesCollectionAI Telemetry Logging (AML.M0024)2
AML.T0006.003 Active Scanning: Probe AI Agent Trigger ChannelsReconnaissanceAI Honeypots (AML.M0039)1T1595
AML.T0024.002 Exfiltration via AI Inference API: Extract AI ModelExfiltrationAI Telemetry Logging (AML.M0024)1
AML.T0043 Craft Adversarial DataAI Attack AdaptationPredictive AI Adversarial Input Detection (AML.M0015)1
AML.T0043.004 Craft Adversarial Data: Insert Backdoor TriggerAI Attack AdaptationPredictive AI Adversarial Input Detection (AML.M0015)1
AML.T0051 LLM Prompt InjectionExecutionAI Telemetry Logging (AML.M0024)1
AML.T0052.000 Phishing: Spearphishing via Social Engineering LLMInitial Access, Lateral MovementDeepfake Detection (AML.M0034)1
AML.T0084 Discover AI Agent ConfigurationDiscoveryAI Honeypots (AML.M0039)1
AML.T0114 AI Service Web InterfaceCommand and ControlAI Telemetry Logging (AML.M0024)1
AML.T0124 Autonomous Attack OrchestrationAI Attack AdaptationAI Agent Scope Drift Detection (AML.M0038)1
AML.T0132 Misconfigured or Publicly Exposed AI ServicesInitial AccessAI Honeypots (AML.M0039)1
AML.T0006.000 Active Scanning: Enumerate Hosted AI ResourcesReconnaissanceAI Honeypots (AML.M0039)0
AML.T0006.001 Active Scanning: Query Platform Metadata APIsReconnaissanceAI Honeypots (AML.M0039)0
AML.T0006.002 Active Scanning: Scan for Exposed AI InfrastructureReconnaissanceAI Honeypots (AML.M0039)0T1595
AML.T0024 Exfiltration via AI Inference APIExfiltrationAI Telemetry Logging (AML.M0024)0
AML.T0024.000 Exfiltration via AI Inference API: Infer Training Data MembershipExfiltrationAI Telemetry Logging (AML.M0024)0
AML.T0024.001 Exfiltration via AI Inference API: Invert AI ModelExfiltrationAI Telemetry Logging (AML.M0024)0
AML.T0052.001 Phishing: Deepfake-Assisted PhishingInitial Access, Lateral MovementDeepfake Detection (AML.M0034)0
AML.T0085 Data from AI ServicesCollectionAI Telemetry Logging (AML.M0024)0
AML.T0133 Discover AI Agent Runtime CapabilitiesDiscoveryAI Honeypots (AML.M0039)0

D3FEND

D3FEND links an attack to a defense through a digital artifact. The attack technique accesses, creates, or modifies an artifact such as a process, a file, or a network flow, and a defensive technique analyzes or hardens the same artifact. D3FEND 1.6.0 links 392 Enterprise and ICS technique ids to 156 defensive techniques: 56 detect, 32 harden, 29 isolate, 14 evict, 12 model, 9 restore, 4 deceive. Once the ids ATT&CK has revoked move to their replacements, 387 of them are techniques in ATT&CK 20.1. A detection program reads the Detect techniques, because each one is a named kind of analysis a hypothesis can start from.

In the coverage map, a technique's detail lists the Detect techniques D3FEND links to it, grouped by the artifact they analyze, and counts the rest by tactic. A parent technique counts its sub-techniques' links and a sub-technique counts its parent's, which is how D3FEND's own site shows them, and each link that comes through another id says which. Mobile techniques have no D3FEND links.

Two examples

For T1003.001, LSASS Memory, the map shows 12 defensive techniques, and 3 of them detect:

Detect techniqueArtifact it analyzesLinked through
D3-PLA Process Lineage AnalysisProcessthe technique itself
D3-PSMD Process Self-Modification DetectionProcessthe technique itself
D3-PSA Process Spawn AnalysisProcessthe technique itself

For T1558.003, Kerberoasting, D3FEND links the sub-technique itself to 10 defensive techniques, all through Network Traffic and RPC Network Traffic. The parent, T1558, adds links through Access Token and Credential, for 21 defensive techniques in all, and 10 of them detect:

Detect techniqueArtifact it analyzesLinked through
D3-APCA Application Protocol Command AnalysisNetwork Trafficthe technique itself
D3-CSPP Client-server Payload ProfilingNetwork Trafficthe technique itself
D3-NTCD Network Traffic Community DeviationNetwork Trafficthe technique itself
D3-NTSA Network Traffic Signature AnalysisNetwork Trafficthe technique itself
D3-PHDURA Per Host Download-Upload Ratio AnalysisNetwork Trafficthe technique itself
D3-PMAD Protocol Metadata Anomaly DetectionNetwork Trafficthe technique itself
D3-RTSD Remote Terminal Session DetectionNetwork Trafficthe technique itself
D3-UGLPA User Geolocation Logon Pattern AnalysisNetwork Trafficthe technique itself
D3-CCSA Credential Compromise Scope AnalysisCredentialT1558
D3-RTA RPC Traffic AnalysisRPC Network Trafficthe technique itself

Differences from D3FEND's own data

The extract reads D3FEND's bulk mappings file, which lists each technique's links on its own, and the coverage map folds parents and sub-techniques when it shows one. On 2026-10-20 the folded counts matched D3FEND's API for 5 of the 5 techniques checked; the outside checks page has the numbers.

D3FEND 1.6.0 maps to ATT&CK 19.0, and this site pins 20.1. ATT&CK has since revoked 22 of the ids D3FEND maps. The extract moves their links to the technique ATT&CK names as each one's replacement, such as T1562.001 to T1685 and T1574.002 to T1574.001. 3 more, T0807, T0853, T1053.004, have no current technique, so the map never shows them. Techniques ATT&CK added after 19.0 have no D3FEND links yet. Of ATT&CK's 7 AI techniques, none has a D3FEND link.

Pinning and updating

tools/atlas_extract.py and tools/d3fend_extract.py each read one pinned release and write atlas-2026.09.json and d3fend-1.6.0.json. Rerun them on each release and review the diff, then rerun tools/d3fend_extract.py --check. In the kit, python tools/attack_pin.py pins the ATLAS ids with ATT&CK's in the technique list lint checks against, and --atlas-only re-pins ATLAS alone.

Licenses

MITRE ATLAS data, copyright The MITRE Corporation, licensed under the Apache License 2.0. Content ATLAS adopts from ATT&CK is also subject to the ATT&CK terms of use. MITRE D3FEND data, copyright The MITRE Corporation, licensed under the MIT License and the D3FEND terms of use, which ask that copies keep this notice. The extracts carry these notices, and this site's CC BY-NC license does not cover them.