129. MITRE ATLAS techniques against AI
ShieldMap · AI security, mapped
Every known way to attack an AI system.
- Tactics
- 16
- Techniques
- 129
- Sub-techniques
- 92
- For LLMs and agents
- 84
- ShieldPi covers
- 25
MITRE ATLAS is the shared catalogue of adversary tactics and techniques against AI, built from real attacks and red-team exercises. On the ring each tactic is a section and each technique a bar, its length the case studies that use it; the matrix below has the same techniques in columns. Each links to the advisories, technique pages and ShieldPi checks that go with it. Names and short descriptions only; the full entries are on ATLAS.
Columns run from reconnaissance to impact. Within a tactic, techniques seen in the most case studies come first. A green dot marks techniques a ShieldPi assessment, runtime check or config check covers.
ReconnaissanceAML.TA0002 · 3
Resource DevelopmentAML.TA0003 · 7
- Develop CapabilitiesAML.T00174 sub20 cases
- Acquire Public AI ArtifactsAML.T00023 sub12 cases
- Obtain CapabilitiesAML.T00165 sub10 cases
- Publish Poisoned AI ArtifactsAML.T01153 sub9 cases
- Replicate AI Model BehaviorAML.T00053 sub9 cases
- Compromise AccountsAML.T01392 cases
- Publish Hallucinated EntitiesAML.T00601 case
AI Attack AdaptationAML.TA0001 · 8
- LLM Prompt CraftingAML.T006528 cases
- Replicate AI Model BehaviorAML.T00053 sub9 cases
- Verify AttackAML.T00429 cases
- Manipulate AI ModelAML.T00184 sub7 cases
- Autonomous Attack-Path AdaptationAML.T01176 cases
- Autonomous AI Agent CommunicationAML.T01182 sub3 cases
- Retrieval Content CraftingAML.T00663 cases
- Autonomous Attack OrchestrationAML.T01241 case
Initial AccessAML.TA0004 · 4
AI Model AccessAML.TA0000 · 4
ExecutionAML.TA0005 · 5
PersistenceAML.TA0006 · 10
- Prompt Infiltration via Public-Facing ApplicationAML.T009315 cases
- Manipulate AI ModelAML.T00184 sub7 cases
- AI Agent Context PoisoningAML.T00802 sub6 cases
- RAG PoisoningAML.T00704 cases
- AI Agent Tool PoisoningAML.T01103 sub3 cases
- Modify AI Agent ConfigurationAML.T00813 cases
- Training Data PoisoningAML.T00203 cases
- LLM Prompt Self-ReplicationAML.T00611 case
- AI Agent Tool Data PoisoningAML.T0099
- Poison AI Agent Learning EnvironmentAML.T0145
Privilege EscalationAML.TA0012 · 2
Defense EvasionAML.TA0007 · 15
- LLM JailbreakAML.T005411 cases
- LLM Prompt ObfuscationAML.T006811 cases
- LLM Trusted Output Components ManipulationAML.T00671 sub4 cases
- Centralized AI Service Request RoutingAML.T01403 cases
- Modify AI Agent ConfigurationAML.T00813 cases
- Delay Execution of LLM InstructionsAML.T00942 cases
- Manipulate User LLM Chat HistoryAML.T00922 cases
- AI Service Request Metadata SanitizationAML.T01411 case
- AI Supply Chain Reputation InflationAML.T01111 case
- AI Supply Chain Rug PullAML.T01091 case
- Corrupt AI ModelAML.T00761 case
- False RAG Entry InjectionAML.T00711 case
- Indicator RemovalAML.T01381 case
- AI Targeted CloakingAML.T0134
- Triggers in Multimodal InputsAML.T0129
Credential AccessAML.TA0013 · 3
DiscoveryAML.TA0008 · 8
- Discover AI Agent ConfigurationAML.T00844 sub4 cases
- Discover AI ArtifactsAML.T00073 cases
- Discover AI Model OutputsAML.T00633 cases
- Discover LLM System InformationAML.T00693 sub3 cases
- Discover LLM HallucinationsAML.T00621 case
- Discover AI Agent Runtime CapabilitiesAML.T0133
- Discover AI Model FamilyAML.T0014
- Recover AI Model InternalsAML.T0144
Lateral MovementAML.TA0015 · 2
CollectionAML.TA0009 · 5
Command and ControlAML.TA0014 · 3
ExfiltrationAML.TA0010 · 4
ImpactAML.TA0011 · 9
- External HarmsAML.T00485 sub36 cases
- Erode AI Model IntegrityAML.T00316 cases
- Machine CompromiseAML.T01122 sub4 cases
- Data Destruction via AI Agent Tool InvocationAML.T01013 cases
- Denial of AI ServiceAML.T00292 cases
- AI Agent Response BiasingAML.T01301 case
- Erode Dataset IntegrityAML.T00591 case
- Cost HarvestingAML.T00343 sub
- Spamming AI System with Chaff DataAML.T0046
Data: MITRE ATLAS 2026.10, Copyright 2021-2026 MITRE, under the Apache License 2.0. ShieldMap shows each entry's name and the opening of its description and links to ATLAS for the rest.